L’essentiel (BLUF) : Who Spoke When splits a recording by speaker: it finds who was talking and when, assembles the lines into a dialogue, and can run live subtitles from your microphone. Everything is computed in your browser — the file is never uploaded and no tokens are spent.
This is speaker diarization: the model finds the boundaries of each line and marks it with a speaker — "Speaker 1", "Speaker 2" and so on, up to four speakers. It is useful anywhere a recording leaves you asking "who said that?": interviews, meetings, podcasts, lectures, voice memos.
With speech recognition enabled, text is added on top: words are assigned to speakers and assembled into a readable dialogue with timestamps. Use it right away, or plug the SRT/VTT subtitles with speaker names into your video.
The second tab is live subtitles: turn on the microphone and phrases appear on screen as you speak. Nothing is recorded or uploaded — the subtitles live only on your page.
Pick an audio or video file (MP3, WAV, M4A, MP4, WebM and more) and press the button. The "Also transcribe the speech" checkbox builds a dialogue with text; with it off, the tool only splits the recording by speakers — faster, without the speech model.
The first run downloads the models once (about 380 MB) and caches them in the browser — the next analyses start right away. The result is a "who and when spoke" timeline and, with recognition, a ready dialogue. Copy it or export TXT, SRT, VTT and JSON.
On the Microphone tab subtitles appear phrase by phrase, like teletext: a finished phrase moves into the list with a timestamp, the next one types in below. Handy for calls, lectures and speech-to-text on the go.
Microphone audio goes only to the model inside your browser. There is no recording and no upload — check the network tab in developer tools if you want to verify it yourself.
The model distinguishes up to four speakers. It works best on recordings with reasonably clean sound and clear turn-taking: interviews and calls split confidently, a noisy hall with interruptions — less so. Long recordings are processed in windows, so minutes of audio do not break the analysis.
Transcription accuracy depends on the language and audio quality. The recognition model is trained on English speech: a Russian dialogue is transcribed too, but with more errors than English. For plain transcription in many languages, the separate "Speech Recognition (free)" tool is right there in the same section.
Everything runs on your device: the file never leaves your computer, no tokens are charged and no account is needed — the tool works for guests. WebGPU speeds things up a lot; if the browser does not offer it, the models run on the CPU through WASM — slower, but it works.
The models are heavy: about a gigabyte of memory is needed for a comfortable run. If the page froze on a weaker device, try a paid cloud model or open this page on a stronger computer.
Correct. The models and their weights load into your browser and the computation happens on your device. The file and the results are never sent anywhere — the tool has no server request with audio at all.
On the first run the browser downloads the models (about 380 MB) and caches them. The next runs load the weights from the cache and start immediately.
Yes. The dialogue exports to TXT, SRT, VTT and JSON, and every subtitle line carries its speaker: "Speaker 1", "Speaker 2" and so on.
Up to four speakers sounding at the same time. If only one person spoke in the recording, everything is assigned to "Speaker 1".
The models run their heavy math on the GPU through WebGPU. If the browser does not support it, the tool automatically falls back to the CPU (WASM) — correct results, noticeably slower.
"Speech Recognition" turns a recording into text and subtitles without splitting by speaker. "Who Spoke When" additionally assigns the lines to speakers and can run live subtitles from the microphone.