AI Voice Transcription

Whisper in the browser, and what drives accuracy

Transcription here runs OpenAI's Whisper — an encoder-decoder model trained on a very large, diverse corpus of audio, which is why it's robust to real-world recordings rather than only clean studio speech. It runs in your browser through transformers.js, so the audio is decoded and transcribed on your device. Whisper is open and documented in OpenAI's research (Robust Speech Recognition via Large-Scale Weak Supervision). For clean English audio it's strong — word error rates commonly in the 5-10% range — but accuracy is a function of the input, not a fixed number.

What actually moves the error rate: audio quality first (a close, clean mic beats a phone across a room every time), then how many people talk over each other, then accent and dialect distance from the training data, then domain vocabulary. Background music, low bitrate, and crosstalk are the usual culprits behind a disappointing transcript. If a clip is short and noisy, pinning the language instead of leaving it on auto-detect often cleans up the result, because the model isn't also spending effort guessing which language it's hearing.

Languages, exports and the practical limits

• Languages — auto-detection is tuned to the main 15 Whisper languages (English, Spanish, French, German, Italian, Portuguese, Dutch, Russian, Polish, Turkish, Chinese, Japanese, Korean, Arabic, Hindi). The model knows more, but those are where the multilingual small model is most reliable.
• Subtitle exports — get timestamped .srt or .vtt for YouTube, Premiere and most editors, or plain .txt when you just need the words. Timestamps are segment-level, not word-perfect, so nudge them in your editor for tight captioning.
• Length — up to roughly an hour transcribes comfortably on a typical laptop; longer files are better split, since the whole job runs in your browser's memory.
• First-run cost — the model weights download once (a few hundred MB) and cache, so the first transcript is slow and every one after starts immediately and offline.

Hallucination is real — always proof against the audio

Whisper has one failure mode you must design around: during silence, music, or noisy passages it can confidently insert text that was never spoken — a stray 'thank you', a repeated phrase, or even a fragment of unrelated boilerplate. This isn't a bug in this tool; it's a documented characteristic of the model, which always produces its best guess at words even when there are none. The implication is non-negotiable for any serious use: read the transcript against the audio before you rely on it, paying special attention to gaps and quiet stretches where invented text loves to appear.

That makes the privacy story more than a nicety. Because everything runs locally, a sensitive recording — a medical consult, a legal deposition, a confidential interview — is transcribed without the audio ever being uploaded to a third party, which is exactly the kind of material you should never hand to a cloud service. But local processing doesn't make the output trustworthy on its own; for legal, medical or journalistic transcripts, treat the AI draft as a starting point a human must verify line by line. If you need the spoken content searchable inside a document, the printed-text equivalent is handled by our OCR tool.