How many speakers can it separate, and do I declare the count first?

Up to 32. The count is inferred automatically, so nothing needs declaring. If you do know how many people were in the room you can state it, which improves separation — with a known count the system will not split one person into two.

Can it identify who each person is by name?

No, and it should not. The output uses anonymous labels like speaker_0 and speaker_1, guaranteeing only that one label means one person. It performs no voiceprint identification. You replace the labels with names once and it applies throughout. That is deliberate: matching voices against identities is biometric processing, and a transcription tool should not do it by default.

What about people talking over each other?

Overlapping speech is the acknowledged hard problem in this field. Turn-taking meetings separate very accurately; where two people speak simultaneously for more than a second or two, attribution can be wrong. In practice, use separate microphones where you can, or re-listen to the disputed passage — the timestamps make that a single click.

Meeting Transcription - It Knows Who Said What

Q: The meeting is confidential. Is the audio retained?

Transcription runs under your account and results live in your history, where you can delete them at any time. Note that the audio is processed by a speech recognition service — for highly sensitive material, confirm this meets your organisation's compliance requirements before uploading.

Transcribe meeting recordings with automatic speaker separation for up to 32 people. Every line carries a speaker label and timing, across 90+ languages.

Last updated: 2026-08-26

Ninety percent of writing minutes is working out who said it

Turning an hour of audio into text is the easy part. The hard part is the wall of subject-less sentences that comes back — you have to re-listen to attribute anything.

Speaker diarisation solves exactly that step: attribution happens during transcription, so the output arrives shaped as a conversation rather than one undifferentiated block.

Real output

A two-person English exchange, uploaded with no configuration. Unedited response:

English12 s

Transcript:So where did we land on the migration date? I'd push it a week. QA hasn't signed off on the rollback path yet. Fine, a week, but I want the rollback tested by Thursday

A second recording that sounds like a real stand-up: two people trading turns, cutting in, then splitting the work. The transcript below is exactly what transcription has to recover.

Voice: 科技大佬 / 知性御姐Chinese43 s

Transcript:好,我们开始。先同步一下上周的进度:支付这块,Stripe 和支付宝都已经在线上跑通了,第一笔真实订单是昨天下午。补充一点,退款流程我们也测过了,资金原路退回大概五到十天。剩下的问题是提现,还差一份补充材料。那这周的目标就两个:把补充材料交上去,然后把素材页的音频全部补齐。谁负责? 素材我来,材料你交。周三下午三点再对一次。

Detected language: eng  (confidence 99.3%)
Speakers detected: 2

speaker_0  @0.16s   So where did we land on the migration date?
speaker_1  @3.56s   I'd push it a week. QA hasn't signed off on
                    the rollback path yet.
speaker_0  @8.54s   Fine, a week, but I want the rollback tested
                    by Thursday

Three turns, all attributed correctly, each with a start time. Nothing was configured — the speaker count was not supplied, and no labelling was done.

Look at the third line: when speaker_0 talks again, the system recognises the same person rather than adding a third. That is the difference between diarisation and simply splitting on pauses.

One honest detail: the source script reads "Fine. A week. But I want…" and came back as "Fine, a week, but I want…". Punctuation is inferred from pauses and is the least stable part of any transcript. It does not matter for subtitles; it matters for a formal record.

Anonymous labels, no voiceprint identification

You get speaker_0 and speaker_1, not names. The guarantee is that one label means one person — the system never asks who that person is.

That is a deliberate limit. Matching voices against identities is biometric processing and should not be switched on by default in a transcription tool. Replacing the labels with names takes ten seconds and applies across the whole transcript.

Where it fits

Setting	What separation buys you	Watch out for
Customer interviews	A clean Q&A record; quotes never misattributed	State the count when there are many participants
Two-host podcasts	Dialogue-shaped output, ready for show notes	The category it handles best
Project standups	Who committed to what, with a timestamp to check	Re-listen to any cross-talk
Formal records	Turn-taking is separated very accurately	Confirm compliance for sensitive material

Take that last row seriously: the audio is processed by a speech recognition service. For highly sensitive meetings, confirm that fits your organisation's requirements before uploading.

Where it breaks

Overlapping speech is the known hard problem. Turn-taking is highly accurate; two people speaking at once for more than a second or two can be misattributed.

Two practical mitigations: separate microphones where possible, or re-listen to the disputed passage — with timestamps that is one click, not a hunt along the playhead.

We put this on the page rather than letting you discover it after uploading.

How to use it

Open the voice studio → transcribe
Upload the recording — audio or video, 200MB total per upload
Wait; an hour typically takes one to two minutes
Replace speaker_0 with real names, then export

Speaker separation needs no configuration, and the language is detected automatically across 90+ languages.

FAQ

How many speakers? Declare the count? Up to 32, inferred automatically. Stating a known count improves separation.

Can it name people? No, and it should not — anonymous labels only, no voiceprint identification.

Cross-talk? Separated, but overlaps beyond a second or two can misattribute. Use separate mics or re-listen.

Confidential meetings? Results live under your account and are deletable, but audio is processed by a recognition service — confirm compliance first.

Start here

Cost: 20 credits per file (up to 200MB per upload), refunded automatically if it fails. The $3.99 trial pack has 150 credits, enough for 7 files, and they never expire.

Upload a meeting recording →