About WhisperX Online
An independent web tool by Answer the Web that makes WhisperX-style transcription usable without installation.
Method
- Speech recognition: OpenAI Whisper (tiny, base, small, large-v3-turbo) as ONNX models from the onnx-community on Hugging Face, run with transformers.js and ONNX Runtime Web.
- Word timestamps: cross-attention alignment of the “timestamped” Whisper exports.
- Speaker detection: pyannote segmentation-3.0 finds speech regions (in sections of ten minutes), WeSpeaker ResNet34 – the embedding model pyannote 3 also uses – computes a voice profile per region, and profiles are grouped by cosine similarity. Each phrase gets the speaker with the largest time overlap.
- Export: segments are formed at speaker changes, pauses over one second and sentence ends; JSON follows the WhisperX output structure.
Privacy
Your audio is processed locally. The server only delivers the web page and the program files; model files come from Hugging Face. Details in the privacy policy.
Not affiliated
WhisperX is developed by Max Bain and contributors (github.com/m-bain/whisperX). This site is not affiliated with them, with OpenAI or with pyannote.
Contact
Email: whisperx@llg.marketing