Find files, editable templates and browser test targets by what you need to make or test. The directory below is cut by format; the two collections under it cut the same library by subject and by workflow.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
