Audio & Video Transcriber + Speaker Diarizer
By Ned Walsch · Last updated:
Turn a recording into a timestamped transcript that shows who said what, when — a call, an interview, a livestream, a leaked voicemail. Whisper runs entirely in your browser. The file is never uploaded, there is no signup, and no API key is needed.
A transcript is a lead, not a court record. Automatic speech recognition mishears names, technical terms, numbers, and anything said over noise or crosstalk. Always spot-check the audio at the timestamps that matter before quoting or acting on them. The speaker labels are an approximate grouping, not a verified identification of who is speaking.
Where the processing happens
The audio is decoded in your browser and transcribed with OpenAI Whisper compiled to run on WebAssembly, with WebGPU acceleration when your device supports it. Nothing is uploaded — you can confirm this in your browser’s network tab. The model weights are fetched once from a public CDN and cached, so repeat use is faster and can work offline.
What the speaker separation really is
True speaker diarization uses a trained neural model to build a voice fingerprint for each person. That is heavy to run in a browser, so this tool uses a lighter approach: it groups segments into speakers using the pauses and acoustic shifts between them. It is genuinely useful for following a two- or three-person conversation, but it can merge two similar voices into one label or split a single speaker across several. For a rigorous same-speaker question, use the dedicated Voice Print Matcher, and treat these labels as a reading aid.
Accuracy and what breaks it
Whisper is strong on clear speech but degrades with background noise, overlapping talkers, heavy accents, distant microphones, and low-bitrate recordings. Numbers, proper nouns, and jargon are the first things it gets wrong. The larger the model you pick, the better the accuracy and the slower the run. Whichever model you use, verify the parts that matter against the original audio.
Responsible use
Recording and transcribing people is governed by consent and wiretap laws that vary by jurisdiction — some places require all parties to consent. Only transcribe recordings you have the right to analyze, and be careful with sensitive material such as private calls or medical information. See our Is OSINT legal? overview.
Is my audio uploaded to a server?
No. Transcription runs entirely in your browser via WebAssembly and WebGPU. The file never leaves your device; the model is downloaded once and cached.
How accurate is the speaker separation?
It is an approximate heuristic based on pauses and acoustic shifts, not a trained diarization model. It helps you follow a small-group conversation but can merge or split speakers. Treat the labels as a guide, not an identification.
What files and lengths work best?
Common audio and video formats work; the audio track is extracted automatically. Files under about ten minutes are comfortable on most machines. WebGPU-capable devices are substantially faster than CPU-only fallback.
Which languages are supported?
Whisper auto-detects and handles many languages. Accuracy is highest for clear speech in widely-spoken languages and drops with noise or overlapping speakers.