Skip to content
AudioLab .tools
Sign in

Voice AI / evidence

Cut-off at the start and end: where the measure and the cost agree

What does it cost a recogniser when the first or last part of the speech is missing, and can AudioLab see that a recording was cut?

The measures involved

4 measures: 4 Validated

Show the measures
01 · The test

What we did

The same 24 English read sentences with the first 100, 250 and 500 ms of the speech cut off (everything before the start of the speech, plus that time, removed with a hard cut and no fade) and with the last 250 ms cut off. Two open recognisers, Whisper small and Whisper large-v3-turbo. The conditions 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.

02 · What came out

The result

What it costs a recogniser: the start matters, the end did not. Cutting 250 ms from the start was clearly worse for both models, and 500 ms more so, almost entirely through words that were not written at all: 9.8% of the reference words were deleted at 500 ms. A cut of 100 ms at the start and one of 250 ms at the end made no clear difference.

Whisper small

Condition WER % Change against clean, points (95% interval) Reading
Clean (reference) 5.0 n/a reference
First 100 ms of speech removed 5.7 +0.7 (−1.1 to +3.0) no clear change
First 250 ms of speech removed 9.8 +4.8 (+2.3 to +7.6) clearly worse
First 500 ms of speech removed 14.6 +9.6 (+7.0 to +12.5) clearly worse
Last 250 ms of speech removed 6.2 +1.1 (−0.5 to +2.8) no clear change

Whisper large-v3-turbo

Condition WER % Change against clean, points (95% interval) Reading
Clean (reference) 3.4 n/a reference
First 100 ms of speech removed 5.0 +1.6 (−0.4 to +3.6) no clear change
First 250 ms of speech removed 8.2 +4.8 (+2.4 to +7.3) clearly worse
First 500 ms of speech removed 13.5 +10.0 (+6.8 to +13.3) clearly worse
Last 250 ms of speech removed 4.3 +0.9 (−0.5 to +3.2) 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: a loud edge. The level at the edge of the file separates a cut start from an uncut one: the area under the curve is 0.96 over the 72 cut starts and 24 uncut ones (medians -29, -27 and -24 dBFS against -59 dBFS), and 0.97 for the end (median -32 dBFS against -57 dBFS). The yes-or-no measures are coarser: "cut off at the start" was true for 13, 18 and 20 of 24 recordings cut 100, 250 and 500 ms into the speech (12, 15 and 18 with the default speech detector), "cut off at the end" for 17 of 24 (and 12 with the default), and neither was ever true for the 24 uncut recordings.

Part of what the yes-or-no measures miss is not the measure. By the test's own rule, the first 20 ms of the file is speech in only 19 of the 24 recordings cut at 100 ms, because a hard cut can land in a pause or a soft consonant. Where it does land in speech, the measure flags most of them: 13 of those 19 at 100 ms and 18 of 19 at 500 ms.

So: this is the one place where the measure and the cost line up. A start that was cut hard shows as a loud edge, and that is the damage a recogniser pays for. Use the level at the edge rather than the yes or no.

03 · The reach

How far this reaches

English read speech cut by hand at known positions, 24 sentences per condition, Whisper small and large-v3-turbo on a CPU. The intervals are wide: a change smaller than about 2 points cannot be told from no change.

04 · The limits

What we have not tested

  • Recordings that were really cut off, by a voice detector that opened late or a button pressed late: ours were cut by hand.
  • Streaming recognisers and voice agents, whose handling of the start of an utterance is different.
  • Other recognisers and other languages.
  • We wrote down in advance that a cut of 250 ms at the end 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.

Check a recording Stress-test your own recogniser The whole path

Other evidence