Use Whisper Diarization for speaker labels and word timestamps. Upload a recording, review who said what, and export your result.
Whisper turns speech into text; it does not identify speakers on its own. Whisper Diarization pairs Whisper Large V3 Turbo with pyannote to group voice turns and label the transcript. Whisper Diarization shows who spoke when, with estimated word timestamps. Review overlapping speech and uncertain labels before using the result.
Hear a 28-second excerpt from the model’s published Replicate example. Follow the original speaker labels as you listen.
Audio waveform visualization · Original automatic captions
Speaker 1
But first, can you tell the story of developing of this model and making the complicated decision of how to release it?Speaker 2
Yeah, sure.Whisper Diarization groups voice turns so you can follow each participant. Listen before naming speakers or correcting turns. Labels are not verified identities; similar voices and overlapping speech can cause mistakes.
Review speaker labelsWhisper Diarization supplies estimated word timestamps. Click an available time to replay that passage and correct unclear words. Editing transcript text clears affected word timing; it does not measure new times.
Compare with WhisperX alignmentSave your Whisper Diarization result as labeled TXT, structured JSON, or SRT and VTT subtitles. Edits and downloads use no extra minutes. Subtitle files contain timed text without automatically adding speaker names.
Explore subtitle exportsAdd an audio or video file to Whisper Diarization. Select its language and detect speakers automatically or set an exact count.
Confirm the duration and charge, then run Whisper Diarization. Your task continues when you leave the page.
Review the transcript while listening. Correct text and speaker names, then export your files.
New to Whisper Diarization? Follow the upload and export guide
Whisper Diarization accepts MP3, WAV, M4A, FLAC, OGG, MP4, MOV and WebM. Only upload recordings you have permission to process.
Recording privacy & retentionGOOD TO KNOW
Whisper Diarization combines speech recognition and speaker grouping to show who spoke when. This page runs thomasmol/whisper-diarization on Replicate, pairing Whisper Large V3 Turbo with pyannote speaker-diarization-community-1. You can review and export the result here.
Whisper Diarization uses Large V3 Turbo word timestamps. WhisperX uses large-v3 with a separate forced-alignment stage. Both provide editable transcripts and speaker labels; review either result against your recording.
No. Whisper Diarization groups voices within the recording using pyannote. Listen and rename labels yourself. It does not verify personal identities or recognize the same person across files.
Whisper Diarization accepts Automatic or an exact count from 1 to 50. A minimum/maximum range is unavailable. Review brief replies and overlapping voices, which can produce incorrect groups.
Whisper Diarization offers English on the free trial. Paid access supports English, Spanish, French, German, Italian, Portuguese, Chinese and Japanese. Transcription keeps the spoken language without translation.
Whisper Diarization here uses private file uploads and keeps the spoken language. Translation, vocabulary prompts and external media URLs are unavailable in this editor, although the upstream API has additional inputs.
Whisper Diarization costs one credit per second, rounded up for the recording. Confirm duration before processing. Verification unlocks a one-time five-minute English trial. Failed tasks return held minutes; Whisper Diarization edits and exports use no extra minutes.
Yes. The Whisper Diarization editor opens with a published Replicate example. Play it, inspect word times, edit text, rename labels and download a copy without a paid task. The example may contain errors; it is not an accuracy benchmark.
Explore the example or upload your own conversation.