Speech Recognition & Diarization

Speech-to-text with speaker labels, running entirely in your browser — local Whisper and speaker-diarization models (one-time download, which starts when the page opens). Your audio never leaves this device.

Speech recognition

Record, then click “Stop recording” to transcribe — or pick an audio or video file. Each new recording or file replaces the current audio, transcript, and speaker names.

Labels who spoke when. When off, new audio skips speaker detection and speaker names are hidden in the transcript and exports. Turning it back on detects speakers in the current audio if needed.

Larger models make fewer mistakes, but they download more data and transcribe more slowly. The selected model starts to load one second after you select it, and is used for the next transcription. Each model downloads only the first time; after that, it loads from the browser cache.

Models

Whisper tiny.en · speech-to-text · ~44 MB · runs locally
pyannote + WeSpeaker · speaker diarization · ~34 MB · runs locally
Ready. Click “Start recording”, or transcribe an audio or video file. The models load in the background.

Transcript

No transcript yet. Record or choose an audio or video file to get started.

Speech-to-text with Whisper tiny.en, base.en, or small.en via Transformers.js. Speaker diarization with diarization-js, a port of pyannote speaker-diarization-community-1 by the pyannote.audio team: pyannote segmentation-3.0 (MIT), WeSpeaker ResNet34 speaker embeddings (pyannote/wespeaker-voxceleb-resnet34-LM), and community-1 PLDA parameters; the embedding and PLDA models are licensed CC BY 4.0 (ONNX artifacts). Models run locally with ONNX Runtime Web; only model files are downloaded, never your audio.