iLoveAudioVideo

Speaker diarization online: label who said what

Free, in-browser speaker labels in 99 transcription languages, with TXT, SRT and VTT exports. Nothing is uploaded. No sign-up or length cap.

  • Nothing uploaded
  • Free, no sign-up
  • No length cap

Turn labels on before choosing a file. First use downloads about 108 MB of speaker models, then caches them. Use Chrome on a computer. Works best when people take turns.

Quick answer: To label who said what, turn on Speaker labels, then drop your audio or video file into the tool above. It transcribes with Whisper and labels voices inside your browser. Download TXT, SRT or VTT with speaker labels and timestamps. First use downloads about 108 MB of extra speaker models, cached for later use. There is no length cap. Works best when people take turns.

Updated

How it works

  1. Turn on speaker labelsEnable labels before choosing your recording to follow who said what.
  2. Choose the spoken language and a recordingWhisper supports 99 languages. The file stays in your browser. Speaker models download only when labels are on.
  3. Check the turns, then saveRead the speaker labels beside the timestamps. Click a timestamp to listen, correct the words, and download TXT, SRT or VTT.

What speaker diarization does

Transcription writes the words. Speaker diarization groups stretches of speech by voice, so you can follow who said what in an interview, conversation or meeting recording. The labels group speech by voice without identifying people by name.

A returning voice keeps its label across the file.

What happens when a voice changes

Labels follow the speaker with the most time in each transcript line. When word timestamps are available, a line can split at a speaker change. Otherwise the tool keeps the recognized words together and marks lines with competing voices as mixed/uncertain. It does not guess where a sentence changes hands by counting letters.

Works best when people take turns. When voices overlap, check the labels and words carefully. Check those moments against the recording before quoting anyone.

Downloads, time and long recordings

The first labelled run downloads about 108 MB of speaker models, in addition to the transcription model. The browser caches them for later runs, unless storage is cleared or evicted. Nothing from your recording is uploaded.

Speaker labelling and transcription run one after the other, so labels take extra time. There is no length cap. Longer recordings still depend on your computer’s memory and speed. Use Chrome on a computer and keep the tab open. If labels fail, the tool still tries to deliver the plain transcript.

Speaker model credits

Speaker segmentation uses onnx-community’s ONNX export of pyannote segmentation 3.0, under the MIT licence. Voice grouping uses Microsoft’s WavLM base-plus-sv, converted and quantized by Xenova, under CC BY-SA 3.0. We use those published weights without further changes. The licence requires attribution and share-alike for adaptations of the model; it does not assign that licence to your recording or transcript.

Full speaker model licence and copyright notices.

Questions

Is speaker diarization free, with no upload?

Yes. Both transcription and speaker labelling run in this browser tab. There is no sign-up, charge or recording upload. Model files download to your device.

Is there a recording length limit?

No. There is no length cap. Processing time depends on the recording and your computer’s memory and speed. Keep the tab open until it finishes.

Does it work in every language?

Whisper transcription supports 99 languages. Check names, technical terms and speaker labels against the recording before sharing the transcript.

Can it separate overlapping speech?

Works best when people take turns. Mixed/uncertain labels flag competing voices. Listen back to overlapping speech and check the words and labels before quoting anyone.

Does it work on phones, Safari or Firefox?

Speaker labels are offered for desktop Chrome. On other browsers or phones, or if a speaker model fails to load, the tool tries to deliver a plain transcript and tells you that labels could not finish.

Do TXT, SRT and VTT keep the speaker labels?

Yes. TXT and SRT include speaker prefixes. VTT uses speaker voice tags. The labels group speech by voice; the tool does not identify people by name.