What we did
The same 24 English read sentences from LibriSpeech test-clean in six new conditions: turned down so the speech sits at an active level of -40, -50 and -60 dBFS, and turned up by 12, 20 and 30 dB and hard-clipped at full scale. In the clipped files a median of 0.2%, 5.3% and 25.4% of the samples are at full scale. Two open recognisers, Whisper small and Whisper large-v3-turbo. The conditions, the levels and the way to read the result were registered on 15 October 2026, before the first transcription.
AudioLab's own measures were read on the same files afterwards, without knowing a recogniser result. That reading was exploratory: it had no pass criteria and it is not a pre-registered test.
The result
What it costs a recogniser: nothing we could show. None of the six conditions made a clear difference for either model, not even with a quarter of the samples clipped (a clearly worse reading needs an interval above zero and a change of at least 2 points). Whisper levels the recording itself, and the clipping here is digital: a hard clip at full scale.
Whisper small
| Condition | WER % | Change against clean, points (95% interval) | Reading |
|---|---|---|---|
| Clean (reference) | 5.0 | n/a | reference |
| Speech at an active level of -40 dBFS | 4.6 | −0.5 (−1.6 to +0.5) | no clear change |
| Speech at an active level of -50 dBFS | 4.3 | −0.7 (−1.7 to +0.0) | no clear change |
| Speech at an active level of -60 dBFS | 3.9 | −1.1 (−2.5 to +0.0) | no clear change |
| +12 dB gain, hard clipped at full scale | 4.8 | −0.2 (−1.3 to +0.7) | no clear change |
| +20 dB gain, hard clipped at full scale | 4.6 | −0.5 (−1.3 to +0.0) | no clear change |
| +30 dB gain, hard clipped at full scale | 5.5 | +0.5 (−1.8 to +3.3) | no clear change |
Whisper large-v3-turbo
| Condition | WER % | Change against clean, points (95% interval) | Reading |
|---|---|---|---|
| Clean (reference) | 3.4 | n/a | reference |
| Speech at an active level of -40 dBFS | 3.9 | +0.5 (+0.0 to +1.3) | no clear change |
| Speech at an active level of -50 dBFS | 3.9 | +0.5 (+0.0 to +1.3) | no clear change |
| Speech at an active level of -60 dBFS | 3.0 | −0.5 (−1.4 to +0.0) | no clear change |
| +12 dB gain, hard clipped at full scale | 3.0 | −0.5 (−1.1 to +0.0) | no clear change |
| +20 dB gain, hard clipped at full scale | 3.2 | −0.2 (−0.7 to +0.0) | no clear change |
| +30 dB gain, hard clipped at full scale | 2.7 | −0.7 (−1.8 to +0.0) | no clear change |
English read speech, 24 sentences per condition, the same sentences in every condition, so each recording is compared with its own clean version. Clearly worse means the 95% interval of the paired change lies above zero and the change is at least 2 points. No clear change does not mean no change: the intervals are wide. The conditions and levels were fixed before the first transcription (pre-registered on 15 October 2026). Set stresstest-core-en-v03-new version 0.3.
What AudioLab reads: the clipping, and the level up to a point. Clipping was found in 22 of 24 recordings at +12 dB, 24 of 24 at +20 dB and 24 of 24 at +30 dB, and in none of the 96 recordings without clipping. The two it did not flag at +12 dB hold no run of clipped samples: one has no sample at full scale and the other only two isolated ones, and the measure counts runs of at least three in a row. Over the 72 clipped files the count follows the true share with a rank correlation of 0.96. Light clipping reads low for the same reason: with up to 1 percent of the samples clipped it counted about 0.7 of the samples at full scale (median). The number of clipped regions is not a measure of how much is clipped: the median was 20 regions with 5 percent of the samples clipped and 12 with 25 percent, because heavy clipping makes the stretches merge.
The peak, the RMS level and the crest factor agree with our own count of the samples (largest difference 0.000 dB). The speech level follows the gain: for a drop of 14 dB it read a change of 14.0 dB and for 24 dB 23.8 dB (median). At -60 dBFS the speech detector keeps only about 9 percent of the speech, and in 2 of the 24 recordings none, so the level is the average of the loudest parts only and reads 5.4 dB too high (median over the 22 recordings that have one).
So: AudioLab sees the clipping and, down to -50 dBFS, the level, and the two recognisers did not mind either. The measures are for finding the damage in your recordings, not for predicting what it costs your recogniser. Most of the audio faults we measure cost these two models nothing; what counts is knowing it for yours, and that is what the stress test is for.
How far this reaches
English read speech, 24 sentences per condition, Whisper small and large-v3-turbo on a CPU, so the intervals are wide: a change smaller than about 2 points cannot be told from no change. The levels are a pure gain on clean recordings, and the clipping is a hard clip at full scale of clean recordings turned up.
What we have not tested
- Streaming recognisers and voice agents that listen live: they may react to the level differently.
- Other recognisers than the two open ones, and other languages.
- Clipping at a microphone or in an analogue stage, soft clipping and compressors.
- Recordings that are quiet because the microphone is far away: more noise comes with them.
- We wrote down in advance that the heaviest clipping would cost a recogniser clearly. It did not.
Sources and next steps
Studies in our register: stress test v0.3, V03-ENG. The status of every measure is on /lenses, and how we test is on /methodology.
Other evidence
- Codecs
- Packet loss
- Echo: in preparation
- Noise and SNR
- A competing voice
- Scene measures and content type
- Speech detection
- Cut-off at the start and end
- Room and reverberation
- Do audio measures predict errors?
- A cleaning step, before and after: in preparation
- Where we were wrong