How to separate speakers in a recording
Transcribe for free →Official pricing sources
Official provider sources for verification. See the dated comparisons below; prices and limits can change.
Two different jobs
Transcribing answers what was said; separating speakers answers who said it. They use different models and can fail independently — a clean transcript can still come back with a messy split.
What makes it easier
Close microphones, one person per channel, little crosstalk and distinct voices. An interview recorded properly is close to the best case.
What makes it harder
Speakerphone recordings, several people at once, similar voices and background noise. A crowded meeting room is close to the worst case.
Say how many people there are
Where you know the number, giving it as an upper limit reins in the most common failure, which is one person being split across several labels.
Then give them names
The labels are generic until you replace them. Names you set are applied when a file is generated, so they appear in the text file, the subtitles, the Word document and the JSON alike.
Separating overlapping and multilingual voices: known limits
Speaker separation is a different step from ASR. Overlapping voices, similar timbres, changing microphones and short interjections can lead to merges, splits and wrong speaker turns. The existing internal 9/10 speaker-count check does not establish correct attribution for every overlapping word. Review handovers before naming speakers. There is no published code-switching diarization benchmark.
Frequently asked questions
Why did it find the wrong number of people?
Voices that sound alike can be merged, and one person recorded across two microphones can be split. Setting an upper limit on speakers before running it is the most effective fix.
Does the language matter?
Speaker grouping relies on voice characteristics, but this does not prove equal quality across languages or code-switching. There is no published mixed-language speaker-attribution benchmark. Transcribed words and detected turns should be checked separately.
Will it work on a phone call or a noisy room?
It will run, but expect a worse result. Distance from the microphone and background noise both make voices harder to tell apart.
Is it available on the free plan?
No, speaker separation is part of Premium. Transcription itself, with timestamps, is on the free plan.
Can I separate speakers when two people talk at once?
The feature can label detected turns, but overlapping words and brief interruptions may be assigned to the wrong person or missed. Verify the critical handovers and consider cleaner source audio.