How do you transcribe audio with multiple speakers?

Choose a tool that returns speaker labels, run the audio through it, and then do the two steps the software cannot do: name each speaker and check every change in the labels against the recording. The labels come from speaker diarization, which separates the voices and numbers them. Our guide to timestamps and speaker labels covers the mechanics in depth, and our overview of transcription covers the whole subject.

Step 1: Choose a tool that returns speaker labels

Not every speech-to-text product returns labels, so check before you upload. The table shows what each vendor's documentation says, as of the pages we read in October 2026.

Live sessions are the trap. A tool that transcribes as people talk may not offer labels at all. For a recording you have already made, use a non-streaming option. Google points to its non-streaming audio transcription endpoint for speaker labels in that case ⁵.

Step 2: Check support for your language

Support for labels can depend on the language. Google's documentation says that speaker diarization is not available for every language, and tells you to "Review the language support page to see if this feature is available for your language" ⁶. Check that page before you choose a product for a recording in another language. Our guide to transcribing audio in other languages covers accuracy across languages.

Step 3: Understand the price model

Some vendors bill diarization on its own. AssemblyAI's pricing page lists speaker diarization at $0.02 per hour of audio ⁷. Multiply the hours of audio you have by the rate the vendor currently lists, and read the pricing page the day you buy, because prices change. Our guide to transcription software explains how to compare tools on price and features.

Step 4: Name each speaker once

The labels are numbers, so the names are your job. Replace each number with the person's name once, using the first place that voice speaks to confirm who it is. Do not assume that the order of speakers matches anything in the recording. The software has no way to know that the first voice is the host.

Step 5: Check every change in the labels

Labels tend to go wrong where the speaker changes, where two voices overlap and where turns are short. Play the audio at each change and confirm the voice. Short replies such as "yes" or "right" are the lines to check first, because they may leave a diarizer very little audio to work with. Mark any overlap you cannot separate, instead of guessing who spoke.

The reason to check is that diarization error is measured, not rare. In the second DIHARD challenge, the DKU-LENOVO team reported a diarization error rate of 27.90 percent on Track 2 ⁸. That is a challenge benchmark, so a clean recording may do better, but the size of the error is the reason a check is worth your time.

Three mistakes make speaker labels less trustworthy

• Trusting the numbers as names. "Speaker 2" is a label, not a person. Verify it.
• Skipping the timestamps. Labels without timestamps are hard to check, because you cannot find the moment a voice changed. Turn on both.
• Comparing scores across papers. Diarization scores depend on the scoring rules, so compare numbers only when the rules match.

For more on the timestamp side of the job, see our question on how often to add timestamps to a transcript. For free tools that give timestamps without labels, see how to transcribe audio with timestamps for free. If your question is specifically about Whisper, see whether Whisper can identify different speakers.