How far you can rely on Patchwell
Measured numbers, what they don't cover, and the promises behind the API. Updated November 7, 2026.
Speaker labels: measured, not claimed
We scored speaker labels word by word on 32 public recordings that come with human "who spoke when" answers: AMI meetings, VoxConverse debates and news, Earnings-21 earnings calls and SIMSAMU conversations, with 1 to 17 speakers and up to 45 minutes long. We tuned on 20 of them and checked on the other 12.
| Recording | Words credited to the right speaker |
|---|---|
| Two-person interviews and podcasts | 99 to 100% |
| Four-person meetings | 98 to 99% |
| News shows and debates (up to 17 speakers) | 94 to 99% |
| Earnings calls (7 to 9 speakers, phone lines) | 92 to 99% |
| All 12 checking recordings together | 97.1% |
Scores count words where exactly one person was talking. Words spoken while two people overlap still get a label, but aren't scored.
Where it goes wrong
- One- and two-word replies ("yeah", "okay") at fast handovers. These are most of the misses.
- A short soundbite from someone who speaks once.
- Someone whose voice changes mid-recording, for example moving from a room mic to a phone line. Their later words can go to someone else, and the system can be confident about it. This is the one failure the shading below can miss.
The shading tells you the real odds
Every word carries spk_p_wrong, the chance its speaker label is wrong. We fitted it on some recordings
and checked it on others it had never seen:
| Shown chance | Share of words | Actually wrong |
|---|---|---|
| Under 10% (no shading) | 98.3% | 2.0% |
| 10 to 30% (light) | 0.9% | 12% |
| 30 to 50% (medium) | 0.2% | 35% |
| Over 50% (dark) | 0.6% | 66% |
Each shaded spot comes with a link that opens your recording at that second, so you can hear it and decide.
Names
A speaker gets a name only when the recording says it ("My name is...", "Our next question comes from..."), and the quote it came from is included. We never identify anyone from their voice.
Safe inside automations
- No double charges: retries with the same
Idempotency-Keyreturn the first response and are charged once. - Failures are free: a file that won't download, a refused research question or an answer we couldn't ground costs nothing.
- Signed webhooks: results can be pushed to you, signed with HMAC-SHA256 and retried for about two hours.
- Stable results: every result carries
schema_version; within version 1, fields are only added. - Safe links: we only fetch public http(s) addresses, never private or internal ones.
Your data
Audio and files are processed in temporary containers and discarded. Listening clips are cut from your link when you ask and are never stored. Results of long jobs are held only until you collect them. Details are in the privacy policy.
Keeping the service up
Every deploy runs about 200 automated tests and a set of live checks against the public API before it counts as done. Account balances are backed up daily. If something looks wrong, write to patchwell.support@gmail.com.