Speech recognition fixtures

ASR: 0123 Clean (WAV)

Synthetic clean tone sequence encoding digits 0123 (zero one two three). Pair with the noisy twin and transcript JSON for ASR evaluation.

File: WAV · Asr · 16000
Use case: ASR testingAudio analysis· Paired fixture

ASR: 0123 Noisy (WAV)

Noisy twin of the digit sequence 0123 at ~7 dB SNR. Score ASR against the shared transcript JSON.

File: WAV · Asr · 16000
Use case: ASR testingAudio analysis· Paired fixture

ASR: 0123 Transcript (JSON)

Ground-truth transcript for the 0123 ASR utterance pair, expected text: “zero one two three”.

File: JSON · Asr
Use case: ASR testingJSON parsing· Conversion set

ASR: 4567 Clean (WAV)

Synthetic clean tone sequence encoding digits 4567 (four five six seven). Pair with the noisy twin and transcript JSON for ASR evaluation.

File: WAV · Asr · 16000
Use case: ASR testingAudio analysis· Paired fixture

ASR: 4567 Noisy (WAV)

Noisy twin of the digit sequence 4567 at ~7 dB SNR. Score ASR against the shared transcript JSON.

File: WAV · Asr · 16000
Use case: ASR testingAudio analysis· Paired fixture

ASR: 4567 Transcript (JSON)

Ground-truth transcript for the 4567 ASR utterance pair, expected text: “four five six seven”.

File: JSON · Asr
Use case: ASR testingJSON parsing· Conversion set

ASR: 89 Clean (WAV)

Synthetic clean tone sequence encoding digits 89 (eight nine). Pair with the noisy twin and transcript JSON for ASR evaluation.

File: WAV · Asr · 16000
Use case: ASR testingAudio analysis· Paired fixture

ASR: 89 Noisy (WAV)

Noisy twin of the digit sequence 89 at ~7 dB SNR. Score ASR against the shared transcript JSON.

File: WAV · Asr · 16000
Use case: ASR testingAudio analysis· Paired fixture

ASR: 89 Transcript (JSON)

Ground-truth transcript for the 89 ASR utterance pair, expected text: “eight nine”.

File: JSON · Asr
Use case: ASR testingJSON parsing· Conversion set

ASR: 1357 Clean (WAV)

Synthetic clean tone sequence encoding digits 1357 (one three five seven). Pair with the noisy twin and transcript JSON for ASR evaluation.

File: WAV · Asr · 16000
Use case: ASR testingAudio analysis· Paired fixture

ASR: 1357 Noisy (WAV)

Noisy twin of the digit sequence 1357 at ~7 dB SNR. Score ASR against the shared transcript JSON.

File: WAV · Asr · 16000
Use case: ASR testingAudio analysis· Paired fixture

ASR: 1357 Transcript (JSON)

Ground-truth transcript for the 1357 ASR utterance pair, expected text: “one three five seven”.

File: JSON · Asr
Use case: ASR testingJSON parsing· Conversion set

ASR: 24680 Clean (WAV)

Synthetic clean tone sequence encoding digits 24680 (two four six eight zero). Pair with the noisy twin and transcript JSON for ASR evaluation.

File: WAV · Asr · 16000
Use case: ASR testingAudio analysis· Paired fixture

ASR: 24680 Noisy (WAV)

Noisy twin of the digit sequence 24680 at ~7 dB SNR. Score ASR against the shared transcript JSON.

File: WAV · Asr · 16000
Use case: ASR testingAudio analysis· Paired fixture

ASR: 24680 Transcript (JSON)

Ground-truth transcript for the 24680 ASR utterance pair, expected text: “two four six eight zero”.

File: JSON · Asr
Use case: ASR testingJSON parsing· Conversion set

ASR: 987654 Clean (WAV)

Synthetic clean tone sequence encoding digits 987654 (nine eight seven six five four). Pair with the noisy twin and transcript JSON for ASR evaluation.

File: WAV · Asr · 16000
Use case: ASR testingAudio analysis· Paired fixture

ASR: 987654 Noisy (WAV)

Noisy twin of the digit sequence 987654 at ~7 dB SNR. Score ASR against the shared transcript JSON.

File: WAV · Asr · 16000
Use case: ASR testingAudio analysis· Paired fixture

ASR: 987654 Transcript (JSON)

Ground-truth transcript for the 987654 ASR utterance pair, expected text: “nine eight seven six five four”.

File: JSON · Asr
Use case: ASR testingJSON parsing· Conversion set

ASR Cue: “one” Clean (WAV)

Short synthetic cue tone for “one”: clean reference for wake-word / digit ASR smoke tests.

File: WAV · Asr · 16000
Use case: ASR testingAudio analysis· Paired fixture

ASR Cue: “one” Noisy (WAV)

Noisy twin of the “one” cue tone (~5 dB SNR).

File: WAV · Asr · 16000
Use case: ASR testingAudio analysis· Paired fixture

ASR Cue: “zero” Clean (WAV)

Short synthetic cue tone for “zero”: clean reference for wake-word / digit ASR smoke tests.

File: WAV · Asr · 16000
Use case: ASR testingAudio analysis· Paired fixture

ASR Cue: “zero” Noisy (WAV)

Noisy twin of the “zero” cue tone (~5 dB SNR).

File: WAV · Asr · 16000
Use case: ASR testingAudio analysis· Paired fixture

ASR Extra: 1111 Clean (WAV)

Extra synthetic digit utterance 1111 (one one one one): clean twin.

File: WAV · Asr · 16000
Use case: ASR testingAudio analysis· Paired fixture

ASR Extra: 1111 Noisy (WAV)

Noisy twin of digit sequence 1111.

File: WAV · Asr · 16000
Use case: ASR testingAudio analysis· Paired fixture

ASR Extra: 1111 Transcript (JSON)

Ground-truth transcript for extra ASR utterance 1111.

File: JSON · Asr
Use case: ASR testingJSON parsing· Conversion set

ASR Extra: 2222 Clean (WAV)

Extra synthetic digit utterance 2222 (two two two two): clean twin.

File: WAV · Asr · 16000
Use case: ASR testingAudio analysis· Paired fixture

ASR Extra: 2222 Noisy (WAV)

Noisy twin of digit sequence 2222.

File: WAV · Asr · 16000
Use case: ASR testingAudio analysis· Paired fixture

ASR Extra: 2222 Transcript (JSON)

Ground-truth transcript for extra ASR utterance 2222.

File: JSON · Asr
Use case: ASR testingJSON parsing· Conversion set

ASR Extra: 3333 Clean (WAV)

Extra synthetic digit utterance 3333 (three three three three): clean twin.

File: WAV · Asr · 16000
Use case: ASR testingAudio analysis· Paired fixture

ASR Extra: 3333 Noisy (WAV)

Noisy twin of digit sequence 3333.

File: WAV · Asr · 16000
Use case: ASR testingAudio analysis· Paired fixture

ASR Extra: 3333 Transcript (JSON)

Ground-truth transcript for extra ASR utterance 3333.

File: JSON · Asr
Use case: ASR testingJSON parsing· Conversion set

ASR Extra: 4444 Clean (WAV)

Extra synthetic digit utterance 4444 (four four four four): clean twin.

File: WAV · Asr · 16000
Use case: ASR testingAudio analysis· Paired fixture

ASR Extra: 4444 Noisy (WAV)

Noisy twin of digit sequence 4444.

File: WAV · Asr · 16000
Use case: ASR testingAudio analysis· Paired fixture

ASR Extra: 4444 Transcript (JSON)

Ground-truth transcript for extra ASR utterance 4444.

File: JSON · Asr
Use case: ASR testingJSON parsing· Conversion set

ASR Extra: 5555 Clean (WAV)

Extra synthetic digit utterance 5555 (five five five five): clean twin.

File: WAV · Asr · 16000
Use case: ASR testingAudio analysis· Paired fixture

ASR Extra: 5555 Noisy (WAV)

Noisy twin of digit sequence 5555.

File: WAV · Asr · 16000
Use case: ASR testingAudio analysis· Paired fixture

ASR Extra: 5555 Transcript (JSON)

Ground-truth transcript for extra ASR utterance 5555.

File: JSON · Asr
Use case: ASR testingJSON parsing· Conversion set

ASR Extra: 6666 Clean (WAV)

Extra synthetic digit utterance 6666 (six six six six): clean twin.

File: WAV · Asr · 16000
Use case: ASR testingAudio analysis· Paired fixture

ASR Extra: 6666 Noisy (WAV)

Noisy twin of digit sequence 6666.

File: WAV · Asr · 16000
Use case: ASR testingAudio analysis· Paired fixture

ASR Extra: 6666 Transcript (JSON)

Ground-truth transcript for extra ASR utterance 6666.

File: JSON · Asr
Use case: ASR testingJSON parsing· Conversion set

ASR Extra: 7777 Clean (WAV)

Extra synthetic digit utterance 7777 (seven seven seven seven): clean twin.

File: WAV · Asr · 16000
Use case: ASR testingAudio analysis· Paired fixture

ASR Extra: 7777 Noisy (WAV)

Noisy twin of digit sequence 7777.

File: WAV · Asr · 16000
Use case: ASR testingAudio analysis· Paired fixture

ASR Extra: 7777 Transcript (JSON)

Ground-truth transcript for extra ASR utterance 7777.

File: JSON · Asr
Use case: ASR testingJSON parsing· Conversion set

ASR Extra: 8888 Clean (WAV)

Extra synthetic digit utterance 8888 (eight eight eight eight): clean twin.

File: WAV · Asr · 16000
Use case: ASR testingAudio analysis· Paired fixture

ASR Extra: 8888 Noisy (WAV)

Noisy twin of digit sequence 8888.

File: WAV · Asr · 16000
Use case: ASR testingAudio analysis· Paired fixture

ASR Extra: 8888 Transcript (JSON)

Ground-truth transcript for extra ASR utterance 8888.

File: JSON · Asr
Use case: ASR testingJSON parsing· Conversion set

ASR Extra: 9999 Clean (WAV)

Extra synthetic digit utterance 9999 (nine nine nine nine): clean twin.

File: WAV · Asr · 16000
Use case: ASR testingAudio analysis· Paired fixture

ASR Extra: 9999 Noisy (WAV)

Noisy twin of digit sequence 9999.

File: WAV · Asr · 16000
Use case: ASR testingAudio analysis· Paired fixture

ASR Extra: 9999 Transcript (JSON)

Ground-truth transcript for extra ASR utterance 9999.

File: JSON · Asr
Use case: ASR testingJSON parsing· Conversion set

ASR Extra: 0000 Clean (WAV)

Extra synthetic digit utterance 0000 (zero zero zero zero): clean twin.

File: WAV · Asr · 16000
Use case: ASR testingAudio analysis· Paired fixture

ASR Extra: 0000 Noisy (WAV)

Noisy twin of digit sequence 0000.

File: WAV · Asr · 16000
Use case: ASR testingAudio analysis· Paired fixture

ASR Extra: 0000 Transcript (JSON)

Ground-truth transcript for extra ASR utterance 0000.

File: JSON · Asr
Use case: ASR testingJSON parsing· Conversion set

DTMF Sequence: 1234

Synthetic DTMF tone sequence for digits “1234”. Pair with the transcript JSON.

File: WAV · Dtmf · 16000
Use case: ASR testingAudio analysis· Paired fixture

DTMF Transcript: 1234 (JSON)

Ground-truth digit transcript for the DTMF sequence 1234.

File: JSON · Dtmf
Use case: ASR testing· Paired fixture

DTMF Sequence: 567890

Synthetic DTMF tone sequence for digits “567890”. Pair with the transcript JSON.

File: WAV · Dtmf · 16000
Use case: ASR testingAudio analysis· Paired fixture

DTMF Transcript: 567890 (JSON)

Ground-truth digit transcript for the DTMF sequence 567890.

File: JSON · Dtmf
Use case: ASR testing· Paired fixture

DTMF Sequence: *9#

Synthetic DTMF tone sequence for digits “*9#”. Pair with the transcript JSON.

File: WAV · Dtmf · 16000
Use case: ASR testingAudio analysis· Paired fixture

DTMF Transcript: *9# (JSON)

Ground-truth digit transcript for the DTMF sequence *9#.

File: JSON · Dtmf
Use case: ASR testing· Paired fixture

DTMF Sequence: 042

Synthetic DTMF tone sequence for digits “042”. Pair with the transcript JSON.

File: WAV · Dtmf · 16000
Use case: ASR testingAudio analysis· Paired fixture

DTMF Transcript: 042 (JSON)

Ground-truth digit transcript for the DTMF sequence 042.

File: JSON · Dtmf
Use case: ASR testing· Paired fixture

DTMF Sequence: 13579

Synthetic DTMF tone sequence for digits “13579”. Pair with the transcript JSON.

File: WAV · Dtmf · 16000
Use case: ASR testingAudio analysis· Paired fixture

DTMF Transcript: 13579 (JSON)

Ground-truth digit transcript for the DTMF sequence 13579.

File: JSON · Dtmf
Use case: ASR testing· Paired fixture

Supplier order call: full spoken take

The complete 22.2 second utterance at 24000 Hz mono FLAC, spoken by the af_sarah voice at 1.0x rate. The five-second excerpt in this group's ladder is cut from this take, so anything needing real duration uses this file and anything comparing damage uses the ladder. Lossless, so it is the archival reference for the whole group.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis· Conversion set

Supplier order call: reference script

Exactly what was spoken, which is what the synthesiser was given: numbers, currency and times are spelled out as words because that is how they were said. This is ground truth by construction - it existed before the audio did - and it is deliberately NOT the same string as the recogniser transcript paired with it.

File: TXT · Speech Ladders
Use case: ASR testing· Paired fixture

Supplier order call: recogniser transcript

What a speech recogniser returned for this take. It normalises: spoken "forty three dollars and eighteen cents" comes back as digits and a currency symbol. Compared with the reference script without normalising first, this disagrees on every number in the utterance and a word error rate computed that way reports a fault that does not exist. That gap is the point of the pair.

File: TXT · Speech Ladders
Use case: ASR testing· Paired fixture

Supplier order call: timed captions

The recogniser's output with timings, 6 cues over the full take. Carries the same normalisation as the plain transcript in this group, so it inherits the same scoring trap, and adds the timing dimension: a cue that starts before the word is spoken is a different defect from a cue with the wrong words in it.

File: SRT · Speech Ladders · 6 cues
Use case: ASR testingMedia accessibility· Conversion set

Supplier order call: plate

A 5 second excerpt of this group's take, reference. The clean five-second excerpt every other file in this ladder is scored against. Delivered as flac at 24000 Hz mono. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Supplier order call: noise snr20

A 5 second excerpt of this group's take, noise. Broadband noise mixed in at a measured signal-to-noise ratio, which is where a denoiser is either working or not. Delivered as flac at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Supplier order call: noise snr10

A 5 second excerpt of this group's take, noise. Broadband noise mixed in at a measured signal-to-noise ratio, which is where a denoiser is either working or not. Delivered as flac at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Supplier order call: noise snr0

A 5 second excerpt of this group's take, noise. Broadband noise mixed in at a measured signal-to-noise ratio, which is where a denoiser is either working or not. Delivered as flac at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Supplier order call: clipped

A 5 second excerpt of this group's take, clipping. Samples driven past a fixed threshold and flattened, the damage a hot microphone does and the one no codec can undo. Delivered as flac at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Supplier order call: dropouts

A 5 second excerpt of this group's take, dropouts. Short spans silenced outright, as a packet-switched call drops them, so a transcriber must decide between a pause and a missing word. Delivered as flac at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Supplier order call: rate 44100

A 5 second excerpt of this group's take, resample. The same audio at a different sample rate, which is the step most speech pipelines get wrong silently by assuming their model's rate. Delivered as flac at 44100 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Supplier order call: rate 16000

A 5 second excerpt of this group's take, resample. The same audio at a different sample rate, which is the step most speech pipelines get wrong silently by assuming their model's rate. Delivered as flac at 16000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Supplier order call: rate 8000

A 5 second excerpt of this group's take, resample. The same audio at a different sample rate, which is the step most speech pipelines get wrong silently by assuming their model's rate. Delivered as flac at 8000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Supplier order call: codec mp3 128

A 5 second excerpt of this group's take, codec. Encoded and decoded through a real speech codec, so the artefacts are the ones the format actually produces rather than a simulation of them. Delivered as mp3 at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: MP3 · Speech Ladders · mp3
Use case: ASR testingAudio analysis+1· Conversion set

Supplier order call: codec mp3 32

A 5 second excerpt of this group's take, codec. Encoded and decoded through a real speech codec, so the artefacts are the ones the format actually produces rather than a simulation of them. Delivered as mp3 at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: MP3 · Speech Ladders · mp3
Use case: ASR testingAudio analysis+1· Conversion set

Supplier order call: codec opus 24

A 5 second excerpt of this group's take, codec. Encoded and decoded through a real speech codec, so the artefacts are the ones the format actually produces rather than a simulation of them. Delivered as opus at 48000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: OPUS · Speech Ladders · opus
Use case: ASR testingAudio analysis+1· Conversion set

Supplier order call: codec gsm 8k

A 5 second excerpt of this group's take, codec. Encoded and decoded through a real speech codec, so the artefacts are the ones the format actually produces rather than a simulation of them. Delivered as gsm at 8000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: GSM · Speech Ladders · gsm
Use case: ASR testingAudio analysis+1· Conversion set

Supplier order call: codec mulaw 8k

A 5 second excerpt of this group's take, codec. Encoded and decoded through a real speech codec, so the artefacts are the ones the format actually produces rather than a simulation of them. Delivered as pcm_mulaw at 8000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: WAV · Speech Ladders · pcm_mulaw
Use case: ASR testingAudio analysis+1· Conversion set

Supplier order call: codec g722 16k

A 5 second excerpt of this group's take, codec. Encoded and decoded through a real speech codec, so the artefacts are the ones the format actually produces rather than a simulation of them. Delivered as adpcm_g722 at 16000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: WAV · Speech Ladders · adpcm_g722
Use case: ASR testingAudio analysis+1· Conversion set

Supplier order call: chain telephone

A 5 second excerpt of this group's take, telephone chain. Band-limited, resampled and companded in one pass: the whole telephone path applied at once rather than one effect at a time. Delivered as pcm_mulaw at 8000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: WAV · Speech Ladders · pcm_mulaw
Use case: ASR testingAudio analysis+1· Conversion set

Invoice dispute: full spoken take

The complete 20.2 second utterance at 24000 Hz mono FLAC, spoken by the am_michael voice at 0.95x rate. The five-second excerpt in this group's ladder is cut from this take, so anything needing real duration uses this file and anything comparing damage uses the ladder. Lossless, so it is the archival reference for the whole group.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis· Conversion set

Invoice dispute: reference script

Exactly what was spoken, which is what the synthesiser was given: numbers, currency and times are spelled out as words because that is how they were said. This is ground truth by construction - it existed before the audio did - and it is deliberately NOT the same string as the recogniser transcript paired with it.

File: TXT · Speech Ladders
Use case: ASR testing· Paired fixture

Invoice dispute: recogniser transcript

What a speech recogniser returned for this take. It normalises: spoken "forty three dollars and eighteen cents" comes back as digits and a currency symbol. Compared with the reference script without normalising first, this disagrees on every number in the utterance and a word error rate computed that way reports a fault that does not exist. That gap is the point of the pair.

File: TXT · Speech Ladders
Use case: ASR testing· Paired fixture

Invoice dispute: timed captions

The recogniser's output with timings, 4 cues over the full take. Carries the same normalisation as the plain transcript in this group, so it inherits the same scoring trap, and adds the timing dimension: a cue that starts before the word is spoken is a different defect from a cue with the wrong words in it.

File: SRT · Speech Ladders · 4 cues
Use case: ASR testingMedia accessibility· Conversion set

Invoice dispute: plate

A 5 second excerpt of this group's take, reference. The clean five-second excerpt every other file in this ladder is scored against. Delivered as flac at 24000 Hz mono. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Invoice dispute: noise snr20

A 5 second excerpt of this group's take, noise. Broadband noise mixed in at a measured signal-to-noise ratio, which is where a denoiser is either working or not. Delivered as flac at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Invoice dispute: noise snr10

A 5 second excerpt of this group's take, noise. Broadband noise mixed in at a measured signal-to-noise ratio, which is where a denoiser is either working or not. Delivered as flac at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Invoice dispute: noise snr0

A 5 second excerpt of this group's take, noise. Broadband noise mixed in at a measured signal-to-noise ratio, which is where a denoiser is either working or not. Delivered as flac at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Invoice dispute: clipped

A 5 second excerpt of this group's take, clipping. Samples driven past a fixed threshold and flattened, the damage a hot microphone does and the one no codec can undo. Delivered as flac at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Invoice dispute: dropouts

A 5 second excerpt of this group's take, dropouts. Short spans silenced outright, as a packet-switched call drops them, so a transcriber must decide between a pause and a missing word. Delivered as flac at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Invoice dispute: rate 44100

A 5 second excerpt of this group's take, resample. The same audio at a different sample rate, which is the step most speech pipelines get wrong silently by assuming their model's rate. Delivered as flac at 44100 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Invoice dispute: rate 16000

A 5 second excerpt of this group's take, resample. The same audio at a different sample rate, which is the step most speech pipelines get wrong silently by assuming their model's rate. Delivered as flac at 16000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Invoice dispute: rate 8000

A 5 second excerpt of this group's take, resample. The same audio at a different sample rate, which is the step most speech pipelines get wrong silently by assuming their model's rate. Delivered as flac at 8000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Invoice dispute: codec mp3 128

A 5 second excerpt of this group's take, codec. Encoded and decoded through a real speech codec, so the artefacts are the ones the format actually produces rather than a simulation of them. Delivered as mp3 at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: MP3 · Speech Ladders · mp3
Use case: ASR testingAudio analysis+1· Conversion set

Invoice dispute: codec mp3 32

A 5 second excerpt of this group's take, codec. Encoded and decoded through a real speech codec, so the artefacts are the ones the format actually produces rather than a simulation of them. Delivered as mp3 at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: MP3 · Speech Ladders · mp3
Use case: ASR testingAudio analysis+1· Conversion set

Invoice dispute: codec opus 24

A 5 second excerpt of this group's take, codec. Encoded and decoded through a real speech codec, so the artefacts are the ones the format actually produces rather than a simulation of them. Delivered as opus at 48000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: OPUS · Speech Ladders · opus
Use case: ASR testingAudio analysis+1· Conversion set

Invoice dispute: codec gsm 8k

A 5 second excerpt of this group's take, codec. Encoded and decoded through a real speech codec, so the artefacts are the ones the format actually produces rather than a simulation of them. Delivered as gsm at 8000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: GSM · Speech Ladders · gsm
Use case: ASR testingAudio analysis+1· Conversion set

Invoice dispute: codec mulaw 8k

A 5 second excerpt of this group's take, codec. Encoded and decoded through a real speech codec, so the artefacts are the ones the format actually produces rather than a simulation of them. Delivered as pcm_mulaw at 8000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: WAV · Speech Ladders · pcm_mulaw
Use case: ASR testingAudio analysis+1· Conversion set

Invoice dispute: codec g722 16k

A 5 second excerpt of this group's take, codec. Encoded and decoded through a real speech codec, so the artefacts are the ones the format actually produces rather than a simulation of them. Delivered as adpcm_g722 at 16000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: WAV · Speech Ladders · adpcm_g722
Use case: ASR testingAudio analysis+1· Conversion set

Invoice dispute: chain telephone

A 5 second excerpt of this group's take, telephone chain. Band-limited, resampled and companded in one pass: the whole telephone path applied at once rather than one effect at a time. Delivered as pcm_mulaw at 8000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: WAV · Speech Ladders · pcm_mulaw
Use case: ASR testingAudio analysis+1· Conversion set

Reservation voicemail: full spoken take

The complete 15.6 second utterance at 24000 Hz mono FLAC, spoken by the bf_emma voice at 1.0x rate. The five-second excerpt in this group's ladder is cut from this take, so anything needing real duration uses this file and anything comparing damage uses the ladder. Lossless, so it is the archival reference for the whole group.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis· Conversion set

Reservation voicemail: reference script

Exactly what was spoken, which is what the synthesiser was given: numbers, currency and times are spelled out as words because that is how they were said. This is ground truth by construction - it existed before the audio did - and it is deliberately NOT the same string as the recogniser transcript paired with it.

File: TXT · Speech Ladders
Use case: ASR testing· Paired fixture

Reservation voicemail: recogniser transcript

What a speech recogniser returned for this take. It normalises: spoken "forty three dollars and eighteen cents" comes back as digits and a currency symbol. Compared with the reference script without normalising first, this disagrees on every number in the utterance and a word error rate computed that way reports a fault that does not exist. That gap is the point of the pair.

File: TXT · Speech Ladders
Use case: ASR testing· Paired fixture

Reservation voicemail: timed captions

The recogniser's output with timings, 5 cues over the full take. Carries the same normalisation as the plain transcript in this group, so it inherits the same scoring trap, and adds the timing dimension: a cue that starts before the word is spoken is a different defect from a cue with the wrong words in it.

File: SRT · Speech Ladders · 5 cues
Use case: ASR testingMedia accessibility· Conversion set

Reservation voicemail: plate

A 5 second excerpt of this group's take, reference. The clean five-second excerpt every other file in this ladder is scored against. Delivered as flac at 24000 Hz mono. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Reservation voicemail: noise snr20

A 5 second excerpt of this group's take, noise. Broadband noise mixed in at a measured signal-to-noise ratio, which is where a denoiser is either working or not. Delivered as flac at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Reservation voicemail: noise snr10

A 5 second excerpt of this group's take, noise. Broadband noise mixed in at a measured signal-to-noise ratio, which is where a denoiser is either working or not. Delivered as flac at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Reservation voicemail: noise snr0

A 5 second excerpt of this group's take, noise. Broadband noise mixed in at a measured signal-to-noise ratio, which is where a denoiser is either working or not. Delivered as flac at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Reservation voicemail: clipped

A 5 second excerpt of this group's take, clipping. Samples driven past a fixed threshold and flattened, the damage a hot microphone does and the one no codec can undo. Delivered as flac at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Reservation voicemail: dropouts

A 5 second excerpt of this group's take, dropouts. Short spans silenced outright, as a packet-switched call drops them, so a transcriber must decide between a pause and a missing word. Delivered as flac at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Reservation voicemail: rate 44100

A 5 second excerpt of this group's take, resample. The same audio at a different sample rate, which is the step most speech pipelines get wrong silently by assuming their model's rate. Delivered as flac at 44100 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Reservation voicemail: rate 16000

A 5 second excerpt of this group's take, resample. The same audio at a different sample rate, which is the step most speech pipelines get wrong silently by assuming their model's rate. Delivered as flac at 16000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Reservation voicemail: rate 8000

A 5 second excerpt of this group's take, resample. The same audio at a different sample rate, which is the step most speech pipelines get wrong silently by assuming their model's rate. Delivered as flac at 8000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Reservation voicemail: codec mp3 128

A 5 second excerpt of this group's take, codec. Encoded and decoded through a real speech codec, so the artefacts are the ones the format actually produces rather than a simulation of them. Delivered as mp3 at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: MP3 · Speech Ladders · mp3
Use case: ASR testingAudio analysis+1· Conversion set

Reservation voicemail: codec mp3 32

A 5 second excerpt of this group's take, codec. Encoded and decoded through a real speech codec, so the artefacts are the ones the format actually produces rather than a simulation of them. Delivered as mp3 at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: MP3 · Speech Ladders · mp3
Use case: ASR testingAudio analysis+1· Conversion set

Reservation voicemail: codec opus 24

A 5 second excerpt of this group's take, codec. Encoded and decoded through a real speech codec, so the artefacts are the ones the format actually produces rather than a simulation of them. Delivered as opus at 48000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: OPUS · Speech Ladders · opus
Use case: ASR testingAudio analysis+1· Conversion set

Reservation voicemail: codec gsm 8k

A 5 second excerpt of this group's take, codec. Encoded and decoded through a real speech codec, so the artefacts are the ones the format actually produces rather than a simulation of them. Delivered as gsm at 8000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: GSM · Speech Ladders · gsm
Use case: ASR testingAudio analysis+1· Conversion set

Reservation voicemail: codec mulaw 8k

A 5 second excerpt of this group's take, codec. Encoded and decoded through a real speech codec, so the artefacts are the ones the format actually produces rather than a simulation of them. Delivered as pcm_mulaw at 8000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: WAV · Speech Ladders · pcm_mulaw
Use case: ASR testingAudio analysis+1· Conversion set

Reservation voicemail: codec g722 16k

A 5 second excerpt of this group's take, codec. Encoded and decoded through a real speech codec, so the artefacts are the ones the format actually produces rather than a simulation of them. Delivered as adpcm_g722 at 16000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: WAV · Speech Ladders · adpcm_g722
Use case: ASR testingAudio analysis+1· Conversion set

Reservation voicemail: chain telephone

A 5 second excerpt of this group's take, telephone chain. Band-limited, resampled and companded in one pass: the whole telephone path applied at once rather than one effect at a time. Delivered as pcm_mulaw at 8000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: WAV · Speech Ladders · pcm_mulaw
Use case: ASR testingAudio analysis+1· Conversion set

Delivery driver: full spoken take

The complete 13.0 second utterance at 24000 Hz mono FLAC, spoken by the am_adam voice at 1.05x rate. The five-second excerpt in this group's ladder is cut from this take, so anything needing real duration uses this file and anything comparing damage uses the ladder. Lossless, so it is the archival reference for the whole group.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis· Conversion set

Delivery driver: reference script

Exactly what was spoken, which is what the synthesiser was given: numbers, currency and times are spelled out as words because that is how they were said. This is ground truth by construction - it existed before the audio did - and it is deliberately NOT the same string as the recogniser transcript paired with it.

File: TXT · Speech Ladders
Use case: ASR testing· Paired fixture

Delivery driver: recogniser transcript

What a speech recogniser returned for this take. It normalises: spoken "forty three dollars and eighteen cents" comes back as digits and a currency symbol. Compared with the reference script without normalising first, this disagrees on every number in the utterance and a word error rate computed that way reports a fault that does not exist. That gap is the point of the pair.

File: TXT · Speech Ladders
Use case: ASR testing· Paired fixture

Delivery driver: timed captions

The recogniser's output with timings, 6 cues over the full take. Carries the same normalisation as the plain transcript in this group, so it inherits the same scoring trap, and adds the timing dimension: a cue that starts before the word is spoken is a different defect from a cue with the wrong words in it.

File: SRT · Speech Ladders · 6 cues
Use case: ASR testingMedia accessibility· Conversion set

Delivery driver: plate

A 5 second excerpt of this group's take, reference. The clean five-second excerpt every other file in this ladder is scored against. Delivered as flac at 24000 Hz mono. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Delivery driver: noise snr20

A 5 second excerpt of this group's take, noise. Broadband noise mixed in at a measured signal-to-noise ratio, which is where a denoiser is either working or not. Delivered as flac at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Delivery driver: noise snr10

A 5 second excerpt of this group's take, noise. Broadband noise mixed in at a measured signal-to-noise ratio, which is where a denoiser is either working or not. Delivered as flac at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Delivery driver: noise snr0

A 5 second excerpt of this group's take, noise. Broadband noise mixed in at a measured signal-to-noise ratio, which is where a denoiser is either working or not. Delivered as flac at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Delivery driver: clipped

A 5 second excerpt of this group's take, clipping. Samples driven past a fixed threshold and flattened, the damage a hot microphone does and the one no codec can undo. Delivered as flac at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Delivery driver: dropouts

A 5 second excerpt of this group's take, dropouts. Short spans silenced outright, as a packet-switched call drops them, so a transcriber must decide between a pause and a missing word. Delivered as flac at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Delivery driver: rate 44100

A 5 second excerpt of this group's take, resample. The same audio at a different sample rate, which is the step most speech pipelines get wrong silently by assuming their model's rate. Delivered as flac at 44100 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Delivery driver: rate 16000

A 5 second excerpt of this group's take, resample. The same audio at a different sample rate, which is the step most speech pipelines get wrong silently by assuming their model's rate. Delivered as flac at 16000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Delivery driver: rate 8000

A 5 second excerpt of this group's take, resample. The same audio at a different sample rate, which is the step most speech pipelines get wrong silently by assuming their model's rate. Delivered as flac at 8000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Delivery driver: codec mp3 128

A 5 second excerpt of this group's take, codec. Encoded and decoded through a real speech codec, so the artefacts are the ones the format actually produces rather than a simulation of them. Delivered as mp3 at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: MP3 · Speech Ladders · mp3
Use case: ASR testingAudio analysis+1· Conversion set

Delivery driver: codec mp3 32

A 5 second excerpt of this group's take, codec. Encoded and decoded through a real speech codec, so the artefacts are the ones the format actually produces rather than a simulation of them. Delivered as mp3 at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: MP3 · Speech Ladders · mp3
Use case: ASR testingAudio analysis+1· Conversion set

Delivery driver: codec opus 24

A 5 second excerpt of this group's take, codec. Encoded and decoded through a real speech codec, so the artefacts are the ones the format actually produces rather than a simulation of them. Delivered as opus at 48000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: OPUS · Speech Ladders · opus
Use case: ASR testingAudio analysis+1· Conversion set

Delivery driver: codec gsm 8k

A 5 second excerpt of this group's take, codec. Encoded and decoded through a real speech codec, so the artefacts are the ones the format actually produces rather than a simulation of them. Delivered as gsm at 8000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: GSM · Speech Ladders · gsm
Use case: ASR testingAudio analysis+1· Conversion set

Delivery driver: codec mulaw 8k

A 5 second excerpt of this group's take, codec. Encoded and decoded through a real speech codec, so the artefacts are the ones the format actually produces rather than a simulation of them. Delivered as pcm_mulaw at 8000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: WAV · Speech Ladders · pcm_mulaw
Use case: ASR testingAudio analysis+1· Conversion set

Delivery driver: codec g722 16k

A 5 second excerpt of this group's take, codec. Encoded and decoded through a real speech codec, so the artefacts are the ones the format actually produces rather than a simulation of them. Delivered as adpcm_g722 at 16000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: WAV · Speech Ladders · adpcm_g722
Use case: ASR testingAudio analysis+1· Conversion set

Delivery driver: chain telephone

A 5 second excerpt of this group's take, telephone chain. Band-limited, resampled and companded in one pass: the whole telephone path applied at once rather than one effect at a time. Delivered as pcm_mulaw at 8000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: WAV · Speech Ladders · pcm_mulaw
Use case: ASR testingAudio analysis+1· Conversion set

Staff briefing: full spoken take

The complete 26.2 second utterance at 24000 Hz mono FLAC, spoken by the af_nicole voice at 1.0x rate. The five-second excerpt in this group's ladder is cut from this take, so anything needing real duration uses this file and anything comparing damage uses the ladder. Lossless, so it is the archival reference for the whole group.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis· Conversion set

Staff briefing: reference script

Exactly what was spoken, which is what the synthesiser was given: numbers, currency and times are spelled out as words because that is how they were said. This is ground truth by construction - it existed before the audio did - and it is deliberately NOT the same string as the recogniser transcript paired with it.

File: TXT · Speech Ladders
Use case: ASR testing· Paired fixture

Staff briefing: recogniser transcript

What a speech recogniser returned for this take. It normalises: spoken "forty three dollars and eighteen cents" comes back as digits and a currency symbol. Compared with the reference script without normalising first, this disagrees on every number in the utterance and a word error rate computed that way reports a fault that does not exist. That gap is the point of the pair.

File: TXT · Speech Ladders
Use case: ASR testing· Paired fixture

Staff briefing: timed captions

The recogniser's output with timings, 6 cues over the full take. Carries the same normalisation as the plain transcript in this group, so it inherits the same scoring trap, and adds the timing dimension: a cue that starts before the word is spoken is a different defect from a cue with the wrong words in it.

File: SRT · Speech Ladders · 6 cues
Use case: ASR testingMedia accessibility· Conversion set

Staff briefing: plate

A 5 second excerpt of this group's take, reference. The clean five-second excerpt every other file in this ladder is scored against. Delivered as flac at 24000 Hz mono. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Staff briefing: noise snr20

A 5 second excerpt of this group's take, noise. Broadband noise mixed in at a measured signal-to-noise ratio, which is where a denoiser is either working or not. Delivered as flac at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Staff briefing: noise snr10

A 5 second excerpt of this group's take, noise. Broadband noise mixed in at a measured signal-to-noise ratio, which is where a denoiser is either working or not. Delivered as flac at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Staff briefing: noise snr0

A 5 second excerpt of this group's take, noise. Broadband noise mixed in at a measured signal-to-noise ratio, which is where a denoiser is either working or not. Delivered as flac at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Staff briefing: clipped

A 5 second excerpt of this group's take, clipping. Samples driven past a fixed threshold and flattened, the damage a hot microphone does and the one no codec can undo. Delivered as flac at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Staff briefing: dropouts

A 5 second excerpt of this group's take, dropouts. Short spans silenced outright, as a packet-switched call drops them, so a transcriber must decide between a pause and a missing word. Delivered as flac at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Staff briefing: rate 44100

A 5 second excerpt of this group's take, resample. The same audio at a different sample rate, which is the step most speech pipelines get wrong silently by assuming their model's rate. Delivered as flac at 44100 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Staff briefing: rate 16000

A 5 second excerpt of this group's take, resample. The same audio at a different sample rate, which is the step most speech pipelines get wrong silently by assuming their model's rate. Delivered as flac at 16000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Staff briefing: rate 8000

A 5 second excerpt of this group's take, resample. The same audio at a different sample rate, which is the step most speech pipelines get wrong silently by assuming their model's rate. Delivered as flac at 8000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Staff briefing: codec mp3 128

A 5 second excerpt of this group's take, codec. Encoded and decoded through a real speech codec, so the artefacts are the ones the format actually produces rather than a simulation of them. Delivered as mp3 at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: MP3 · Speech Ladders · mp3
Use case: ASR testingAudio analysis+1· Conversion set

Staff briefing: codec mp3 32

A 5 second excerpt of this group's take, codec. Encoded and decoded through a real speech codec, so the artefacts are the ones the format actually produces rather than a simulation of them. Delivered as mp3 at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: MP3 · Speech Ladders · mp3
Use case: ASR testingAudio analysis+1· Conversion set

Staff briefing: codec opus 24

A 5 second excerpt of this group's take, codec. Encoded and decoded through a real speech codec, so the artefacts are the ones the format actually produces rather than a simulation of them. Delivered as opus at 48000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: OPUS · Speech Ladders · opus
Use case: ASR testingAudio analysis+1· Conversion set

Staff briefing: codec gsm 8k

A 5 second excerpt of this group's take, codec. Encoded and decoded through a real speech codec, so the artefacts are the ones the format actually produces rather than a simulation of them. Delivered as gsm at 8000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: GSM · Speech Ladders · gsm
Use case: ASR testingAudio analysis+1· Conversion set

Staff briefing: codec mulaw 8k

A 5 second excerpt of this group's take, codec. Encoded and decoded through a real speech codec, so the artefacts are the ones the format actually produces rather than a simulation of them. Delivered as pcm_mulaw at 8000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: WAV · Speech Ladders · pcm_mulaw
Use case: ASR testingAudio analysis+1· Conversion set

Staff briefing: codec g722 16k

A 5 second excerpt of this group's take, codec. Encoded and decoded through a real speech codec, so the artefacts are the ones the format actually produces rather than a simulation of them. Delivered as adpcm_g722 at 16000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: WAV · Speech Ladders · adpcm_g722
Use case: ASR testingAudio analysis+1· Conversion set

Staff briefing: chain telephone

A 5 second excerpt of this group's take, telephone chain. Band-limited, resampled and companded in one pass: the whole telephone path applied at once rather than one effect at a time. Delivered as pcm_mulaw at 8000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: WAV · Speech Ladders · pcm_mulaw
Use case: ASR testingAudio analysis+1· Conversion set

Haccp training: full spoken take

The complete 19.2 second utterance at 24000 Hz mono FLAC, spoken by the bm_george voice at 0.9x rate. The five-second excerpt in this group's ladder is cut from this take, so anything needing real duration uses this file and anything comparing damage uses the ladder. Lossless, so it is the archival reference for the whole group.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis· Conversion set

Haccp training: reference script

Exactly what was spoken, which is what the synthesiser was given: numbers, currency and times are spelled out as words because that is how they were said. This is ground truth by construction - it existed before the audio did - and it is deliberately NOT the same string as the recogniser transcript paired with it.

File: TXT · Speech Ladders
Use case: ASR testing· Paired fixture

Haccp training: recogniser transcript

What a speech recogniser returned for this take. It normalises: spoken "forty three dollars and eighteen cents" comes back as digits and a currency symbol. Compared with the reference script without normalising first, this disagrees on every number in the utterance and a word error rate computed that way reports a fault that does not exist. That gap is the point of the pair.

File: TXT · Speech Ladders
Use case: ASR testing· Paired fixture

Haccp training: timed captions

The recogniser's output with timings, 4 cues over the full take. Carries the same normalisation as the plain transcript in this group, so it inherits the same scoring trap, and adds the timing dimension: a cue that starts before the word is spoken is a different defect from a cue with the wrong words in it.

File: SRT · Speech Ladders · 4 cues
Use case: ASR testingMedia accessibility· Conversion set

Haccp training: plate

A 5 second excerpt of this group's take, reference. The clean five-second excerpt every other file in this ladder is scored against. Delivered as flac at 24000 Hz mono. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Haccp training: noise snr20

A 5 second excerpt of this group's take, noise. Broadband noise mixed in at a measured signal-to-noise ratio, which is where a denoiser is either working or not. Delivered as flac at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Haccp training: noise snr10

A 5 second excerpt of this group's take, noise. Broadband noise mixed in at a measured signal-to-noise ratio, which is where a denoiser is either working or not. Delivered as flac at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Haccp training: noise snr0

A 5 second excerpt of this group's take, noise. Broadband noise mixed in at a measured signal-to-noise ratio, which is where a denoiser is either working or not. Delivered as flac at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Haccp training: clipped

A 5 second excerpt of this group's take, clipping. Samples driven past a fixed threshold and flattened, the damage a hot microphone does and the one no codec can undo. Delivered as flac at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Haccp training: dropouts

A 5 second excerpt of this group's take, dropouts. Short spans silenced outright, as a packet-switched call drops them, so a transcriber must decide between a pause and a missing word. Delivered as flac at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Haccp training: rate 44100

A 5 second excerpt of this group's take, resample. The same audio at a different sample rate, which is the step most speech pipelines get wrong silently by assuming their model's rate. Delivered as flac at 44100 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Haccp training: rate 16000

A 5 second excerpt of this group's take, resample. The same audio at a different sample rate, which is the step most speech pipelines get wrong silently by assuming their model's rate. Delivered as flac at 16000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Haccp training: rate 8000

A 5 second excerpt of this group's take, resample. The same audio at a different sample rate, which is the step most speech pipelines get wrong silently by assuming their model's rate. Delivered as flac at 8000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Haccp training: codec mp3 128

A 5 second excerpt of this group's take, codec. Encoded and decoded through a real speech codec, so the artefacts are the ones the format actually produces rather than a simulation of them. Delivered as mp3 at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: MP3 · Speech Ladders · mp3
Use case: ASR testingAudio analysis+1· Conversion set

Haccp training: codec mp3 32

A 5 second excerpt of this group's take, codec. Encoded and decoded through a real speech codec, so the artefacts are the ones the format actually produces rather than a simulation of them. Delivered as mp3 at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: MP3 · Speech Ladders · mp3
Use case: ASR testingAudio analysis+1· Conversion set

Haccp training: codec opus 24

A 5 second excerpt of this group's take, codec. Encoded and decoded through a real speech codec, so the artefacts are the ones the format actually produces rather than a simulation of them. Delivered as opus at 48000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: OPUS · Speech Ladders · opus
Use case: ASR testingAudio analysis+1· Conversion set

Haccp training: codec gsm 8k

A 5 second excerpt of this group's take, codec. Encoded and decoded through a real speech codec, so the artefacts are the ones the format actually produces rather than a simulation of them. Delivered as gsm at 8000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: GSM · Speech Ladders · gsm
Use case: ASR testingAudio analysis+1· Conversion set

Haccp training: codec mulaw 8k

A 5 second excerpt of this group's take, codec. Encoded and decoded through a real speech codec, so the artefacts are the ones the format actually produces rather than a simulation of them. Delivered as pcm_mulaw at 8000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: WAV · Speech Ladders · pcm_mulaw
Use case: ASR testingAudio analysis+1· Conversion set

Haccp training: codec g722 16k

A 5 second excerpt of this group's take, codec. Encoded and decoded through a real speech codec, so the artefacts are the ones the format actually produces rather than a simulation of them. Delivered as adpcm_g722 at 16000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: WAV · Speech Ladders · adpcm_g722
Use case: ASR testingAudio analysis+1· Conversion set

Haccp training: chain telephone

A 5 second excerpt of this group's take, telephone chain. Band-limited, resampled and companded in one pass: the whole telephone path applied at once rather than one effect at a time. Delivered as pcm_mulaw at 8000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: WAV · Speech Ladders · pcm_mulaw
Use case: ASR testingAudio analysis+1· Conversion set

Ivr menu: full spoken take

The complete 11.9 second utterance at 24000 Hz mono FLAC, spoken by the af_heart voice at 1.0x rate. The five-second excerpt in this group's ladder is cut from this take, so anything needing real duration uses this file and anything comparing damage uses the ladder. Lossless, so it is the archival reference for the whole group.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis· Conversion set

Ivr menu: reference script

Exactly what was spoken, which is what the synthesiser was given: numbers, currency and times are spelled out as words because that is how they were said. This is ground truth by construction - it existed before the audio did - and it is deliberately NOT the same string as the recogniser transcript paired with it.

File: TXT · Speech Ladders
Use case: ASR testing· Paired fixture

Ivr menu: recogniser transcript

What a speech recogniser returned for this take. It normalises: spoken "forty three dollars and eighteen cents" comes back as digits and a currency symbol. Compared with the reference script without normalising first, this disagrees on every number in the utterance and a word error rate computed that way reports a fault that does not exist. That gap is the point of the pair.

File: TXT · Speech Ladders
Use case: ASR testing· Paired fixture

Ivr menu: timed captions

The recogniser's output with timings, 3 cues over the full take. Carries the same normalisation as the plain transcript in this group, so it inherits the same scoring trap, and adds the timing dimension: a cue that starts before the word is spoken is a different defect from a cue with the wrong words in it.

File: SRT · Speech Ladders · 3 cues
Use case: ASR testingMedia accessibility· Conversion set

Ivr menu: plate

A 5 second excerpt of this group's take, reference. The clean five-second excerpt every other file in this ladder is scored against. Delivered as flac at 24000 Hz mono. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Ivr menu: noise snr20

A 5 second excerpt of this group's take, noise. Broadband noise mixed in at a measured signal-to-noise ratio, which is where a denoiser is either working or not. Delivered as flac at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Ivr menu: noise snr10

A 5 second excerpt of this group's take, noise. Broadband noise mixed in at a measured signal-to-noise ratio, which is where a denoiser is either working or not. Delivered as flac at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Ivr menu: noise snr0

A 5 second excerpt of this group's take, noise. Broadband noise mixed in at a measured signal-to-noise ratio, which is where a denoiser is either working or not. Delivered as flac at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Ivr menu: clipped

A 5 second excerpt of this group's take, clipping. Samples driven past a fixed threshold and flattened, the damage a hot microphone does and the one no codec can undo. Delivered as flac at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Ivr menu: dropouts

A 5 second excerpt of this group's take, dropouts. Short spans silenced outright, as a packet-switched call drops them, so a transcriber must decide between a pause and a missing word. Delivered as flac at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Ivr menu: rate 44100

A 5 second excerpt of this group's take, resample. The same audio at a different sample rate, which is the step most speech pipelines get wrong silently by assuming their model's rate. Delivered as flac at 44100 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Ivr menu: rate 16000

A 5 second excerpt of this group's take, resample. The same audio at a different sample rate, which is the step most speech pipelines get wrong silently by assuming their model's rate. Delivered as flac at 16000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Ivr menu: rate 8000

A 5 second excerpt of this group's take, resample. The same audio at a different sample rate, which is the step most speech pipelines get wrong silently by assuming their model's rate. Delivered as flac at 8000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Ivr menu: codec mp3 128

A 5 second excerpt of this group's take, codec. Encoded and decoded through a real speech codec, so the artefacts are the ones the format actually produces rather than a simulation of them. Delivered as mp3 at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: MP3 · Speech Ladders · mp3
Use case: ASR testingAudio analysis+1· Conversion set

Ivr menu: codec mp3 32

A 5 second excerpt of this group's take, codec. Encoded and decoded through a real speech codec, so the artefacts are the ones the format actually produces rather than a simulation of them. Delivered as mp3 at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: MP3 · Speech Ladders · mp3
Use case: ASR testingAudio analysis+1· Conversion set

Ivr menu: codec opus 24

A 5 second excerpt of this group's take, codec. Encoded and decoded through a real speech codec, so the artefacts are the ones the format actually produces rather than a simulation of them. Delivered as opus at 48000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: OPUS · Speech Ladders · opus
Use case: ASR testingAudio analysis+1· Conversion set

Ivr menu: codec gsm 8k

A 5 second excerpt of this group's take, codec. Encoded and decoded through a real speech codec, so the artefacts are the ones the format actually produces rather than a simulation of them. Delivered as gsm at 8000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: GSM · Speech Ladders · gsm
Use case: ASR testingAudio analysis+1· Conversion set

Ivr menu: codec mulaw 8k

A 5 second excerpt of this group's take, codec. Encoded and decoded through a real speech codec, so the artefacts are the ones the format actually produces rather than a simulation of them. Delivered as pcm_mulaw at 8000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: WAV · Speech Ladders · pcm_mulaw
Use case: ASR testingAudio analysis+1· Conversion set

Ivr menu: codec g722 16k

A 5 second excerpt of this group's take, codec. Encoded and decoded through a real speech codec, so the artefacts are the ones the format actually produces rather than a simulation of them. Delivered as adpcm_g722 at 16000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: WAV · Speech Ladders · adpcm_g722
Use case: ASR testingAudio analysis+1· Conversion set

Ivr menu: chain telephone

A 5 second excerpt of this group's take, telephone chain. Band-limited, resampled and companded in one pass: the whole telephone path applied at once rather than one effect at a time. Delivered as pcm_mulaw at 8000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: WAV · Speech Ladders · pcm_mulaw
Use case: ASR testingAudio analysis+1· Conversion set

Cost podcast: full spoken take

The complete 16.5 second utterance at 24000 Hz mono FLAC, spoken by the am_eric voice at 1.0x rate. The five-second excerpt in this group's ladder is cut from this take, so anything needing real duration uses this file and anything comparing damage uses the ladder. Lossless, so it is the archival reference for the whole group.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis· Conversion set

Cost podcast: reference script

Exactly what was spoken, which is what the synthesiser was given: numbers, currency and times are spelled out as words because that is how they were said. This is ground truth by construction - it existed before the audio did - and it is deliberately NOT the same string as the recogniser transcript paired with it.

File: TXT · Speech Ladders
Use case: ASR testing· Paired fixture

Cost podcast: recogniser transcript

What a speech recogniser returned for this take. It normalises: spoken "forty three dollars and eighteen cents" comes back as digits and a currency symbol. Compared with the reference script without normalising first, this disagrees on every number in the utterance and a word error rate computed that way reports a fault that does not exist. That gap is the point of the pair.

File: TXT · Speech Ladders
Use case: ASR testing· Paired fixture

Cost podcast: timed captions

The recogniser's output with timings, 4 cues over the full take. Carries the same normalisation as the plain transcript in this group, so it inherits the same scoring trap, and adds the timing dimension: a cue that starts before the word is spoken is a different defect from a cue with the wrong words in it.

File: SRT · Speech Ladders · 4 cues
Use case: ASR testingMedia accessibility· Conversion set

Cost podcast: plate

A 5 second excerpt of this group's take, reference. The clean five-second excerpt every other file in this ladder is scored against. Delivered as flac at 24000 Hz mono. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Cost podcast: noise snr20

A 5 second excerpt of this group's take, noise. Broadband noise mixed in at a measured signal-to-noise ratio, which is where a denoiser is either working or not. Delivered as flac at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Cost podcast: noise snr10

A 5 second excerpt of this group's take, noise. Broadband noise mixed in at a measured signal-to-noise ratio, which is where a denoiser is either working or not. Delivered as flac at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Cost podcast: noise snr0

A 5 second excerpt of this group's take, noise. Broadband noise mixed in at a measured signal-to-noise ratio, which is where a denoiser is either working or not. Delivered as flac at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Cost podcast: clipped

A 5 second excerpt of this group's take, clipping. Samples driven past a fixed threshold and flattened, the damage a hot microphone does and the one no codec can undo. Delivered as flac at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Cost podcast: dropouts

A 5 second excerpt of this group's take, dropouts. Short spans silenced outright, as a packet-switched call drops them, so a transcriber must decide between a pause and a missing word. Delivered as flac at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Cost podcast: rate 44100

A 5 second excerpt of this group's take, resample. The same audio at a different sample rate, which is the step most speech pipelines get wrong silently by assuming their model's rate. Delivered as flac at 44100 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Cost podcast: rate 16000

A 5 second excerpt of this group's take, resample. The same audio at a different sample rate, which is the step most speech pipelines get wrong silently by assuming their model's rate. Delivered as flac at 16000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Cost podcast: rate 8000

A 5 second excerpt of this group's take, resample. The same audio at a different sample rate, which is the step most speech pipelines get wrong silently by assuming their model's rate. Delivered as flac at 8000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Cost podcast: codec mp3 128

A 5 second excerpt of this group's take, codec. Encoded and decoded through a real speech codec, so the artefacts are the ones the format actually produces rather than a simulation of them. Delivered as mp3 at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: MP3 · Speech Ladders · mp3
Use case: ASR testingAudio analysis+1· Conversion set

Cost podcast: codec mp3 32

A 5 second excerpt of this group's take, codec. Encoded and decoded through a real speech codec, so the artefacts are the ones the format actually produces rather than a simulation of them. Delivered as mp3 at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: MP3 · Speech Ladders · mp3
Use case: ASR testingAudio analysis+1· Conversion set

Cost podcast: codec opus 24

A 5 second excerpt of this group's take, codec. Encoded and decoded through a real speech codec, so the artefacts are the ones the format actually produces rather than a simulation of them. Delivered as opus at 48000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: OPUS · Speech Ladders · opus
Use case: ASR testingAudio analysis+1· Conversion set

Cost podcast: codec gsm 8k

A 5 second excerpt of this group's take, codec. Encoded and decoded through a real speech codec, so the artefacts are the ones the format actually produces rather than a simulation of them. Delivered as gsm at 8000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: GSM · Speech Ladders · gsm
Use case: ASR testingAudio analysis+1· Conversion set

Cost podcast: codec mulaw 8k

A 5 second excerpt of this group's take, codec. Encoded and decoded through a real speech codec, so the artefacts are the ones the format actually produces rather than a simulation of them. Delivered as pcm_mulaw at 8000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: WAV · Speech Ladders · pcm_mulaw
Use case: ASR testingAudio analysis+1· Conversion set

Cost podcast: codec g722 16k

A 5 second excerpt of this group's take, codec. Encoded and decoded through a real speech codec, so the artefacts are the ones the format actually produces rather than a simulation of them. Delivered as adpcm_g722 at 16000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: WAV · Speech Ladders · adpcm_g722
Use case: ASR testingAudio analysis+1· Conversion set

Cost podcast: chain telephone

A 5 second excerpt of this group's take, telephone chain. Band-limited, resampled and companded in one pass: the whole telephone path applied at once rather than one effect at a time. Delivered as pcm_mulaw at 8000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: WAV · Speech Ladders · pcm_mulaw
Use case: ASR testingAudio analysis+1· Conversion set

Complaint call: full spoken take

The complete 13.5 second utterance at 24000 Hz mono FLAC, spoken by the bf_isabella voice at 0.95x rate. The five-second excerpt in this group's ladder is cut from this take, so anything needing real duration uses this file and anything comparing damage uses the ladder. Lossless, so it is the archival reference for the whole group.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis· Conversion set

Complaint call: reference script

Exactly what was spoken, which is what the synthesiser was given: numbers, currency and times are spelled out as words because that is how they were said. This is ground truth by construction - it existed before the audio did - and it is deliberately NOT the same string as the recogniser transcript paired with it.

File: TXT · Speech Ladders
Use case: ASR testing· Paired fixture

Complaint call: recogniser transcript

What a speech recogniser returned for this take. It normalises: spoken "forty three dollars and eighteen cents" comes back as digits and a currency symbol. Compared with the reference script without normalising first, this disagrees on every number in the utterance and a word error rate computed that way reports a fault that does not exist. That gap is the point of the pair.

File: TXT · Speech Ladders
Use case: ASR testing· Paired fixture

Complaint call: timed captions

The recogniser's output with timings, 4 cues over the full take. Carries the same normalisation as the plain transcript in this group, so it inherits the same scoring trap, and adds the timing dimension: a cue that starts before the word is spoken is a different defect from a cue with the wrong words in it.

File: SRT · Speech Ladders · 4 cues
Use case: ASR testingMedia accessibility· Conversion set

Complaint call: plate

A 5 second excerpt of this group's take, reference. The clean five-second excerpt every other file in this ladder is scored against. Delivered as flac at 24000 Hz mono. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Complaint call: noise snr20

A 5 second excerpt of this group's take, noise. Broadband noise mixed in at a measured signal-to-noise ratio, which is where a denoiser is either working or not. Delivered as flac at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Complaint call: noise snr10

A 5 second excerpt of this group's take, noise. Broadband noise mixed in at a measured signal-to-noise ratio, which is where a denoiser is either working or not. Delivered as flac at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Complaint call: noise snr0

A 5 second excerpt of this group's take, noise. Broadband noise mixed in at a measured signal-to-noise ratio, which is where a denoiser is either working or not. Delivered as flac at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Complaint call: clipped

A 5 second excerpt of this group's take, clipping. Samples driven past a fixed threshold and flattened, the damage a hot microphone does and the one no codec can undo. Delivered as flac at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Complaint call: dropouts

A 5 second excerpt of this group's take, dropouts. Short spans silenced outright, as a packet-switched call drops them, so a transcriber must decide between a pause and a missing word. Delivered as flac at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Complaint call: rate 44100

A 5 second excerpt of this group's take, resample. The same audio at a different sample rate, which is the step most speech pipelines get wrong silently by assuming their model's rate. Delivered as flac at 44100 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Complaint call: rate 16000

A 5 second excerpt of this group's take, resample. The same audio at a different sample rate, which is the step most speech pipelines get wrong silently by assuming their model's rate. Delivered as flac at 16000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Complaint call: rate 8000

A 5 second excerpt of this group's take, resample. The same audio at a different sample rate, which is the step most speech pipelines get wrong silently by assuming their model's rate. Delivered as flac at 8000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Complaint call: codec mp3 128

A 5 second excerpt of this group's take, codec. Encoded and decoded through a real speech codec, so the artefacts are the ones the format actually produces rather than a simulation of them. Delivered as mp3 at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: MP3 · Speech Ladders · mp3
Use case: ASR testingAudio analysis+1· Conversion set

Complaint call: codec mp3 32

A 5 second excerpt of this group's take, codec. Encoded and decoded through a real speech codec, so the artefacts are the ones the format actually produces rather than a simulation of them. Delivered as mp3 at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: MP3 · Speech Ladders · mp3
Use case: ASR testingAudio analysis+1· Conversion set

Complaint call: codec opus 24

A 5 second excerpt of this group's take, codec. Encoded and decoded through a real speech codec, so the artefacts are the ones the format actually produces rather than a simulation of them. Delivered as opus at 48000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: OPUS · Speech Ladders · opus
Use case: ASR testingAudio analysis+1· Conversion set

Complaint call: codec gsm 8k

A 5 second excerpt of this group's take, codec. Encoded and decoded through a real speech codec, so the artefacts are the ones the format actually produces rather than a simulation of them. Delivered as gsm at 8000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: GSM · Speech Ladders · gsm
Use case: ASR testingAudio analysis+1· Conversion set

Complaint call: codec mulaw 8k

A 5 second excerpt of this group's take, codec. Encoded and decoded through a real speech codec, so the artefacts are the ones the format actually produces rather than a simulation of them. Delivered as pcm_mulaw at 8000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: WAV · Speech Ladders · pcm_mulaw
Use case: ASR testingAudio analysis+1· Conversion set

Complaint call: codec g722 16k

A 5 second excerpt of this group's take, codec. Encoded and decoded through a real speech codec, so the artefacts are the ones the format actually produces rather than a simulation of them. Delivered as adpcm_g722 at 16000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: WAV · Speech Ladders · adpcm_g722
Use case: ASR testingAudio analysis+1· Conversion set

Complaint call: chain telephone

A 5 second excerpt of this group's take, telephone chain. Band-limited, resampled and companded in one pass: the whole telephone path applied at once rather than one effect at a time. Delivered as pcm_mulaw at 8000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: WAV · Speech Ladders · pcm_mulaw
Use case: ASR testingAudio analysis+1· Conversion set

Numbers stress: full spoken take

The complete 18.5 second utterance at 24000 Hz mono FLAC, spoken by the am_liam voice at 1.0x rate. The five-second excerpt in this group's ladder is cut from this take, so anything needing real duration uses this file and anything comparing damage uses the ladder. Lossless, so it is the archival reference for the whole group.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis· Conversion set

Numbers stress: reference script

Exactly what was spoken, which is what the synthesiser was given: numbers, currency and times are spelled out as words because that is how they were said. This is ground truth by construction - it existed before the audio did - and it is deliberately NOT the same string as the recogniser transcript paired with it.

File: TXT · Speech Ladders
Use case: ASR testing· Paired fixture

Numbers stress: recogniser transcript

What a speech recogniser returned for this take. It normalises: spoken "forty three dollars and eighteen cents" comes back as digits and a currency symbol. Compared with the reference script without normalising first, this disagrees on every number in the utterance and a word error rate computed that way reports a fault that does not exist. That gap is the point of the pair.

File: TXT · Speech Ladders
Use case: ASR testing· Paired fixture

Numbers stress: timed captions

The recogniser's output with timings, 4 cues over the full take. Carries the same normalisation as the plain transcript in this group, so it inherits the same scoring trap, and adds the timing dimension: a cue that starts before the word is spoken is a different defect from a cue with the wrong words in it.

File: SRT · Speech Ladders · 4 cues
Use case: ASR testingMedia accessibility· Conversion set

Numbers stress: plate

A 5 second excerpt of this group's take, reference. The clean five-second excerpt every other file in this ladder is scored against. Delivered as flac at 24000 Hz mono. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Numbers stress: noise snr20

A 5 second excerpt of this group's take, noise. Broadband noise mixed in at a measured signal-to-noise ratio, which is where a denoiser is either working or not. Delivered as flac at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Numbers stress: noise snr10

A 5 second excerpt of this group's take, noise. Broadband noise mixed in at a measured signal-to-noise ratio, which is where a denoiser is either working or not. Delivered as flac at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Numbers stress: noise snr0

A 5 second excerpt of this group's take, noise. Broadband noise mixed in at a measured signal-to-noise ratio, which is where a denoiser is either working or not. Delivered as flac at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Numbers stress: clipped

A 5 second excerpt of this group's take, clipping. Samples driven past a fixed threshold and flattened, the damage a hot microphone does and the one no codec can undo. Delivered as flac at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Numbers stress: dropouts

A 5 second excerpt of this group's take, dropouts. Short spans silenced outright, as a packet-switched call drops them, so a transcriber must decide between a pause and a missing word. Delivered as flac at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Numbers stress: rate 44100

A 5 second excerpt of this group's take, resample. The same audio at a different sample rate, which is the step most speech pipelines get wrong silently by assuming their model's rate. Delivered as flac at 44100 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Numbers stress: rate 16000

A 5 second excerpt of this group's take, resample. The same audio at a different sample rate, which is the step most speech pipelines get wrong silently by assuming their model's rate. Delivered as flac at 16000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Numbers stress: rate 8000

A 5 second excerpt of this group's take, resample. The same audio at a different sample rate, which is the step most speech pipelines get wrong silently by assuming their model's rate. Delivered as flac at 8000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Numbers stress: codec mp3 128

A 5 second excerpt of this group's take, codec. Encoded and decoded through a real speech codec, so the artefacts are the ones the format actually produces rather than a simulation of them. Delivered as mp3 at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: MP3 · Speech Ladders · mp3
Use case: ASR testingAudio analysis+1· Conversion set

Numbers stress: codec mp3 32

A 5 second excerpt of this group's take, codec. Encoded and decoded through a real speech codec, so the artefacts are the ones the format actually produces rather than a simulation of them. Delivered as mp3 at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: MP3 · Speech Ladders · mp3
Use case: ASR testingAudio analysis+1· Conversion set

Numbers stress: codec opus 24

A 5 second excerpt of this group's take, codec. Encoded and decoded through a real speech codec, so the artefacts are the ones the format actually produces rather than a simulation of them. Delivered as opus at 48000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: OPUS · Speech Ladders · opus
Use case: ASR testingAudio analysis+1· Conversion set

Numbers stress: codec gsm 8k

A 5 second excerpt of this group's take, codec. Encoded and decoded through a real speech codec, so the artefacts are the ones the format actually produces rather than a simulation of them. Delivered as gsm at 8000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: GSM · Speech Ladders · gsm
Use case: ASR testingAudio analysis+1· Conversion set

Numbers stress: codec mulaw 8k

A 5 second excerpt of this group's take, codec. Encoded and decoded through a real speech codec, so the artefacts are the ones the format actually produces rather than a simulation of them. Delivered as pcm_mulaw at 8000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: WAV · Speech Ladders · pcm_mulaw
Use case: ASR testingAudio analysis+1· Conversion set

Numbers stress: codec g722 16k

A 5 second excerpt of this group's take, codec. Encoded and decoded through a real speech codec, so the artefacts are the ones the format actually produces rather than a simulation of them. Delivered as adpcm_g722 at 16000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: WAV · Speech Ladders · adpcm_g722
Use case: ASR testingAudio analysis+1· Conversion set

Numbers stress: chain telephone

A 5 second excerpt of this group's take, telephone chain. Band-limited, resampled and companded in one pass: the whole telephone path applied at once rather than one effect at a time. Delivered as pcm_mulaw at 8000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: WAV · Speech Ladders · pcm_mulaw
Use case: ASR testingAudio analysis+1· Conversion set

Homophone stress: full spoken take

The complete 13.3 second utterance at 24000 Hz mono FLAC, spoken by the bm_lewis voice at 1.0x rate. The five-second excerpt in this group's ladder is cut from this take, so anything needing real duration uses this file and anything comparing damage uses the ladder. Lossless, so it is the archival reference for the whole group.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis· Conversion set

Homophone stress: reference script

Exactly what was spoken, which is what the synthesiser was given: numbers, currency and times are spelled out as words because that is how they were said. This is ground truth by construction - it existed before the audio did - and it is deliberately NOT the same string as the recogniser transcript paired with it.

File: TXT · Speech Ladders
Use case: ASR testing· Paired fixture

Homophone stress: recogniser transcript

What a speech recogniser returned for this take. It normalises: spoken "forty three dollars and eighteen cents" comes back as digits and a currency symbol. Compared with the reference script without normalising first, this disagrees on every number in the utterance and a word error rate computed that way reports a fault that does not exist. That gap is the point of the pair.

File: TXT · Speech Ladders
Use case: ASR testing· Paired fixture

Homophone stress: timed captions

The recogniser's output with timings, 5 cues over the full take. Carries the same normalisation as the plain transcript in this group, so it inherits the same scoring trap, and adds the timing dimension: a cue that starts before the word is spoken is a different defect from a cue with the wrong words in it.

File: SRT · Speech Ladders · 5 cues
Use case: ASR testingMedia accessibility· Conversion set

Homophone stress: plate

A 5 second excerpt of this group's take, reference. The clean five-second excerpt every other file in this ladder is scored against. Delivered as flac at 24000 Hz mono. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Homophone stress: noise snr20

A 5 second excerpt of this group's take, noise. Broadband noise mixed in at a measured signal-to-noise ratio, which is where a denoiser is either working or not. Delivered as flac at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Homophone stress: noise snr10

A 5 second excerpt of this group's take, noise. Broadband noise mixed in at a measured signal-to-noise ratio, which is where a denoiser is either working or not. Delivered as flac at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Homophone stress: noise snr0

A 5 second excerpt of this group's take, noise. Broadband noise mixed in at a measured signal-to-noise ratio, which is where a denoiser is either working or not. Delivered as flac at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Homophone stress: clipped

A 5 second excerpt of this group's take, clipping. Samples driven past a fixed threshold and flattened, the damage a hot microphone does and the one no codec can undo. Delivered as flac at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Homophone stress: dropouts

A 5 second excerpt of this group's take, dropouts. Short spans silenced outright, as a packet-switched call drops them, so a transcriber must decide between a pause and a missing word. Delivered as flac at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Homophone stress: rate 44100

A 5 second excerpt of this group's take, resample. The same audio at a different sample rate, which is the step most speech pipelines get wrong silently by assuming their model's rate. Delivered as flac at 44100 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Homophone stress: rate 16000

A 5 second excerpt of this group's take, resample. The same audio at a different sample rate, which is the step most speech pipelines get wrong silently by assuming their model's rate. Delivered as flac at 16000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Homophone stress: rate 8000

A 5 second excerpt of this group's take, resample. The same audio at a different sample rate, which is the step most speech pipelines get wrong silently by assuming their model's rate. Delivered as flac at 8000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Homophone stress: codec mp3 128

A 5 second excerpt of this group's take, codec. Encoded and decoded through a real speech codec, so the artefacts are the ones the format actually produces rather than a simulation of them. Delivered as mp3 at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: MP3 · Speech Ladders · mp3
Use case: ASR testingAudio analysis+1· Conversion set

Homophone stress: codec mp3 32

A 5 second excerpt of this group's take, codec. Encoded and decoded through a real speech codec, so the artefacts are the ones the format actually produces rather than a simulation of them. Delivered as mp3 at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: MP3 · Speech Ladders · mp3
Use case: ASR testingAudio analysis+1· Conversion set

Homophone stress: codec opus 24

A 5 second excerpt of this group's take, codec. Encoded and decoded through a real speech codec, so the artefacts are the ones the format actually produces rather than a simulation of them. Delivered as opus at 48000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: OPUS · Speech Ladders · opus
Use case: ASR testingAudio analysis+1· Conversion set

Homophone stress: codec gsm 8k

A 5 second excerpt of this group's take, codec. Encoded and decoded through a real speech codec, so the artefacts are the ones the format actually produces rather than a simulation of them. Delivered as gsm at 8000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: GSM · Speech Ladders · gsm
Use case: ASR testingAudio analysis+1· Conversion set

Homophone stress: codec mulaw 8k

A 5 second excerpt of this group's take, codec. Encoded and decoded through a real speech codec, so the artefacts are the ones the format actually produces rather than a simulation of them. Delivered as pcm_mulaw at 8000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: WAV · Speech Ladders · pcm_mulaw
Use case: ASR testingAudio analysis+1· Conversion set

Homophone stress: codec g722 16k

A 5 second excerpt of this group's take, codec. Encoded and decoded through a real speech codec, so the artefacts are the ones the format actually produces rather than a simulation of them. Delivered as adpcm_g722 at 16000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: WAV · Speech Ladders · adpcm_g722
Use case: ASR testingAudio analysis+1· Conversion set

Homophone stress: chain telephone

A 5 second excerpt of this group's take, telephone chain. Band-limited, resampled and companded in one pass: the whole telephone path applied at once rather than one effect at a time. Delivered as pcm_mulaw at 8000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: WAV · Speech Ladders · pcm_mulaw
Use case: ASR testingAudio analysis+1· Conversion set

Menu narration: full spoken take

The complete 15.6 second utterance at 24000 Hz mono FLAC, spoken by the af_bella voice at 0.95x rate. The five-second excerpt in this group's ladder is cut from this take, so anything needing real duration uses this file and anything comparing damage uses the ladder. Lossless, so it is the archival reference for the whole group.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis· Conversion set

Menu narration: reference script

Exactly what was spoken, which is what the synthesiser was given: numbers, currency and times are spelled out as words because that is how they were said. This is ground truth by construction - it existed before the audio did - and it is deliberately NOT the same string as the recogniser transcript paired with it.

File: TXT · Speech Ladders
Use case: ASR testing· Paired fixture

Menu narration: recogniser transcript

What a speech recogniser returned for this take. It normalises: spoken "forty three dollars and eighteen cents" comes back as digits and a currency symbol. Compared with the reference script without normalising first, this disagrees on every number in the utterance and a word error rate computed that way reports a fault that does not exist. That gap is the point of the pair.

File: TXT · Speech Ladders
Use case: ASR testing· Paired fixture

Menu narration: timed captions

The recogniser's output with timings, 4 cues over the full take. Carries the same normalisation as the plain transcript in this group, so it inherits the same scoring trap, and adds the timing dimension: a cue that starts before the word is spoken is a different defect from a cue with the wrong words in it.

File: SRT · Speech Ladders · 4 cues
Use case: ASR testingMedia accessibility· Conversion set

Menu narration: plate

A 5 second excerpt of this group's take, reference. The clean five-second excerpt every other file in this ladder is scored against. Delivered as flac at 24000 Hz mono. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Menu narration: noise snr20

A 5 second excerpt of this group's take, noise. Broadband noise mixed in at a measured signal-to-noise ratio, which is where a denoiser is either working or not. Delivered as flac at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Menu narration: noise snr10

A 5 second excerpt of this group's take, noise. Broadband noise mixed in at a measured signal-to-noise ratio, which is where a denoiser is either working or not. Delivered as flac at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Menu narration: noise snr0

A 5 second excerpt of this group's take, noise. Broadband noise mixed in at a measured signal-to-noise ratio, which is where a denoiser is either working or not. Delivered as flac at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Menu narration: clipped

A 5 second excerpt of this group's take, clipping. Samples driven past a fixed threshold and flattened, the damage a hot microphone does and the one no codec can undo. Delivered as flac at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Menu narration: dropouts

A 5 second excerpt of this group's take, dropouts. Short spans silenced outright, as a packet-switched call drops them, so a transcriber must decide between a pause and a missing word. Delivered as flac at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Menu narration: rate 44100

A 5 second excerpt of this group's take, resample. The same audio at a different sample rate, which is the step most speech pipelines get wrong silently by assuming their model's rate. Delivered as flac at 44100 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Menu narration: rate 16000

A 5 second excerpt of this group's take, resample. The same audio at a different sample rate, which is the step most speech pipelines get wrong silently by assuming their model's rate. Delivered as flac at 16000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Menu narration: rate 8000

A 5 second excerpt of this group's take, resample. The same audio at a different sample rate, which is the step most speech pipelines get wrong silently by assuming their model's rate. Delivered as flac at 8000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: FLAC · Speech Ladders · flac
Use case: ASR testingAudio analysis+1· Conversion set

Menu narration: codec mp3 128

A 5 second excerpt of this group's take, codec. Encoded and decoded through a real speech codec, so the artefacts are the ones the format actually produces rather than a simulation of them. Delivered as mp3 at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: MP3 · Speech Ladders · mp3
Use case: ASR testingAudio analysis+1· Conversion set

Menu narration: codec mp3 32

A 5 second excerpt of this group's take, codec. Encoded and decoded through a real speech codec, so the artefacts are the ones the format actually produces rather than a simulation of them. Delivered as mp3 at 24000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: MP3 · Speech Ladders · mp3
Use case: ASR testingAudio analysis+1· Conversion set

Menu narration: codec opus 24

A 5 second excerpt of this group's take, codec. Encoded and decoded through a real speech codec, so the artefacts are the ones the format actually produces rather than a simulation of them. Delivered as opus at 48000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: OPUS · Speech Ladders · opus
Use case: ASR testingAudio analysis+1· Conversion set

Menu narration: codec gsm 8k

A 5 second excerpt of this group's take, codec. Encoded and decoded through a real speech codec, so the artefacts are the ones the format actually produces rather than a simulation of them. Delivered as gsm at 8000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: GSM · Speech Ladders · gsm
Use case: ASR testingAudio analysis+1· Conversion set

Menu narration: codec mulaw 8k

A 5 second excerpt of this group's take, codec. Encoded and decoded through a real speech codec, so the artefacts are the ones the format actually produces rather than a simulation of them. Delivered as pcm_mulaw at 8000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: WAV · Speech Ladders · pcm_mulaw
Use case: ASR testingAudio analysis+1· Conversion set

Menu narration: codec g722 16k

A 5 second excerpt of this group's take, codec. Encoded and decoded through a real speech codec, so the artefacts are the ones the format actually produces rather than a simulation of them. Delivered as adpcm_g722 at 16000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: WAV · Speech Ladders · adpcm_g722
Use case: ASR testingAudio analysis+1· Conversion set

Menu narration: chain telephone

A 5 second excerpt of this group's take, telephone chain. Band-limited, resampled and companded in one pass: the whole telephone path applied at once rather than one effect at a time. Delivered as pcm_mulaw at 8000 Hz mono. Scored against plate.flac in the same group, which is the same five seconds undamaged. The parameters in the specification are what the file MEASURED, not what was requested of it.

File: WAV · Speech Ladders · pcm_mulaw
Use case: ASR testingAudio analysis+1· Conversion set