How much of a recording is speech
The median recording is 80% speech and one in 6 is less than half. Measured by voice-activity detection on 2,511 real recordings.
The short answer
The median recording people bring to WhipScribe is 80% speech; one in 6 is less than half speech, and across 2,257 hours of audio 706 hours were silence, music or noise.
Production record, 28 July 2026 to 25 October 2026. Quote with the date.
How much of a recording is speech
| Speech in the recording | Recordings | Share |
|---|---|---|
| under 10% | 169 | 7% |
| 10–25% | 59 | 2% |
| 25–50% | 188 | 8% |
| 50–75% | 588 | 23% |
| 75–90% | 1,007 | 40% |
| 90% or more | 500 | 20% |
2,511 finished recordings with a speech measurement. 20% are 90% speech or more; 7% are under 10% speech, which is almost always music, a silent camera file or a recording left running.
Long recordings carry more silence
| Recording length | Recordings | Median speech share |
|---|---|---|
| under 5 min | 554 | 79% |
| 5–15 min | 415 | 78% |
| 15–30 min | 324 | 84% |
| 30–60 min | 489 | 84% |
| 1–2 h | 410 | 79% |
| 2–4 h | 224 | 68% |
| 4 h or more | 95 | 68% |
Recordings of two hours or more are 68% speech (2–4 h) and 68% speech (4 h or more) at the median: breaks, set-up time, applause and the minutes before and after a meeting. This is why WhipScribe starts a free preview at the first speech instead of at second zero.
By uploaded format
| Uploaded format | Recordings | Median speech share | Under half speech |
|---|---|---|---|
| M4A | 548 | 78% | 12% |
| MP3 | 437 | 80% | 17% |
| MP4 | 199 | 82% | 22% |
| WAV | 60 | 78% | 25% |
| MOV | 23 | 75% | 35% |
| OGG | 21 | 80% | 19% |
MOV (35%) and WAV (25%) were most often under half speech.
Method
Source: the WhipScribe production job record, every job created from 28 July 2026 to 25 October 2026 (90 days, UTC). 7,417 jobs were created; 1,856 were removed as internal, test or direct API traffic (a few integrations submit in batches and would weigh one account like hundreds), and each account counts for at most its first 20 jobs (262 jobs over that cap left out). That leaves 5,299 jobs from 4,107 accounts, 3,469 of them finished recordings of a minute or more. Speech share is the fraction of the recording's seconds a voice-activity detector marked as speech, stored with each job.
Privacy: aggregates only. Every row shown stands on at least 20 jobs from at least 5 different accounts; smaller groups are folded into "other" or left out. No name, e-mail, file name, link or transcript text was read into the figures. Computed by scripts/research_production.py, measured 26 October 2026.
Questions
How much of a typical recording is actual speech?
The median is 80%, measured on 2,511 recordings from 28 July 2026 to 25 October 2026.
Why does the share of speech matter?
Silence and music still take upload time and minutes on services that bill by file length. It also explains odd transcripts: a recognizer given long silence or music can produce repeated or invented lines, which WhipScribe's clean-up step removes from the transcript.
How is speech measured?
A voice-activity detector marks which seconds of the audio contain a human voice; the speech share is those seconds divided by the recording's length.
Cite this
WhipScribe, "How much of a recording is speech? Measured on 2,511 files", production data 28 July 2026 to 25 October 2026, measured 26 October 2026. https://whipscribe.com/proof/speech-share
Free to quote and chart with that line (CC BY 4.0). Raw figures: /proof/speech-share.json.
More from the production record: how long recordings are · languages people transcribe · when people transcribe · formats, links and devices · how fast it transcribes · all proof pages · fact sheet.
See also: Transcription services · Turnaround time · Evidence log · Pricing · Privacy · Terms