Try it now
A working sandbox. No sign-up, no project. Sample data only.
- Voice fingerprinting runs on our servers. No raw audio is sent to a third-party service.
- Cross-file matching: name a speaker on Day 04 and Day 17 picks up the label too.
- Merge, split or rename identities project-wide. Renames apply everywhere instantly.
- Correct a turn and the nearby matches are reconsidered.
- Too few speakers found? Re-diarize listens to the audio again and keeps every word, edit and note.
- A change history: every naming, merge and reset, who did it, across how many files.
How it works
Three steps from raw material to result.
Each speaker turn is analysed locally on our server to create a unique voice fingerprint. No audio is sent to a third-party speaker service.
Fingerprints are compared against existing project identities. Strong matches auto-group; ambiguous ones surface for review.
Rename a speaker on any file and the change ripples to every appearance, including files you uploaded weeks ago.
Listen again
When a file comes back with too few speakers, re-diarize relabels the lines in place. The words and everything built on them stay.
Frequently asked questions
How does cross-file matching work?
Each speaker turn is analysed to create a unique voice fingerprint, then compared against the identities already in the project. Matches above a tuned confidence threshold are grouped automatically; ambiguous ones are flagged for manual review.
Can I merge two speakers I named separately?
Yes. Merge from the speaker panel and all appearances of the absorbed identity are renamed across every file. Splitting one identity into two is also supported.
What if the system gets a speaker wrong?
Correct the label on any turn and the system re-analyses and reconsiders nearby matches automatically. A project-wide cleanup pass can be triggered manually after large edits.
Does it work for two speakers who sound alike?
Voice fingerprinting separates most adult voices reliably even when timbre is close. Identical-twin-level similarity occasionally needs manual review, especially in short turns under two seconds.
Is speaker identification run locally?
No. It runs on our servers, not in your browser. No raw audio is sent to a third-party speaker service.
The file came back with one speaker and the room had four. What do I do?
Use Re-diarize on the file. It listens to the audio again with the newest speaker model and keeps your transcript text. A file over ten minutes with one voice throughout gets a warning with the button on it. The optional Expected speakers box is an upper limit that guides the model when it is unsure, and the limit from the first run is not carried over.
What is the difference between Re-diarize and Reset speakers?
Reset speakers rewinds every line to the labels the provider first gave, which can never produce more speakers than the first run found. Re-diarize listens again and can. After a Re-diarize, Reset rewinds to the new split.
Can I tell the same person apart in two places?
Yes. Give the speaker a one-letter location tag in each folder, such as L for London and B for Berkeley. The speaker stays one identity and every name-based feature still treats them as one person. The tag shows as a second, coloured half of the chip in the paper cut, Tara MacAulay | L, and is written in brackets in exports, Tara MacAulay (L). Most speakers will never need one.
Related capabilities
Further reading
Background guides and comparisons.
Diarization assigns speaker labels to audio segments without knowing who the speakers are. Here is how voice prints work, why similar vocal profiles cause problems, and how merge thresholds control the output.
Running your own pyannote diarization stack makes sense for custom fine-tuning, offline requirements, or research. For production documentary work, the real cost is not the model but the annotation and tuning loop.