Every speaker, separated
Speaker diarization that holds. One mixed track goes in, each voice comes out with its own label and lane, and keeps it from the first minute to the last. Up to 25 speakers in one recording.
Every voice, on the record.
Speech recognition for real conversations. Every speaker separated, overlapping speech kept, every word timed.
The lowest errors of any system we tested. See the results
board-review.m4a
0:00 / 0:18- 0:00Speaker 1
So what did you actually tell the board?
- 0:04Speaker 2
That we should hold the launch a week.
- 0:07Speaker 1
A week? We don't have a week.
- 0:09Speaker 3
We do if the second region waits.
- 0:13Speaker 4
Then let's bring both options on Monday.
26%
The lowest errors on every benchmark.
fewer errors than the next best system, on average across the public meeting benchmarks.
- NOTSOFAR office meetings
- 43%
- fewer errors than AssemblyAI and pyannote
- Deepgram Nova-3
- 54–73%
- fewer errors, depending on the benchmark
- AMI headset meetings
- Best
- published result, ahead of Gemini and VibeVoice
The whole conversation, not just the words.
Crosstalk, kept
When two people talk at once, most systems keep one and drop the other. Both are transcribed here, and the overlap is marked with its exact start and end.
Timed to the word
Every word carries its own timestamp, so you can search a recording, jump to a moment, or cut a clip at the exact word.
A crowded room, sorted by voice.
Up to 25 speakers are told apart by the sound of their voice, so the same person keeps the same label, even after long silences.
Add what the transcript needs.
Ten add-ons, switched on per request. They work in all 45 languages, named or detected.
Personal details, removed.
Before
Send it to [email protected] and call me on +1 415 555 0134 after four.
After
Send it to and call me on after four.
The card ending is the one on file.
3 details removed · text, word timings and exports
Your recordings stay yours.
Delete anytime
Remove any file and its transcript whenever you want. Deletion is immediate and permanent.
Private to your workspace
Recordings and results are visible only to your workspace and are never used to train models.
Kept for playback
Audio is kept 7 days so you can play it next to the transcript. Results are kept 30 days.
Calico
Our speech model, built for recordings with more than one voice. It runs every transcript, in the dashboard and through the API.
- Languages
- 45
- Speakers
- 25
- Longest file
- 3 h
- Largest file
- 2 GB
Built into your product in an afternoon.
Upload a file, ask for a transcript, and get the whole conversation back as JSON. A REST API with an OpenAPI contract and signed webhooks.
