You recorded a one-hour meeting and now need the text to search through it, quote it, or archive it. Uploading that recording to a cloud service to transcribe it means sending a potentially confidential conversation to a third party. Transcribing with AI directly in the browser sidesteps that problem entirely: the audio never leaves your device.
What Whisper is and why it runs in the browser
Whisper is OpenAI's speech recognition model, trained on hundreds of thousands of hours of audio across dozens of languages. The version this tool uses is exported to ONNX and runs with transformers.js, which lets inference run on WebGPU (if your browser supports it) or fall back to WebAssembly — no server, no API key, no minute limits.
The three model tiers
- Fast (~75 MB, Whisper tiny): the lightest, good for short clips or when you prioritize speed over accuracy.
- Precise (~150 MB, Whisper base): the recommended default balance for most recordings.
- Max quality (~1 GB, large-v3-turbo): the large model quantized to 8-bit, OpenAI's pick for the best accuracy-to-size ratio as of 2026. Worth it for strong accents, noisy background audio, or when accuracy matters more than wait time.
The model downloads once and stays cached in the browser — the second transcription at the same tier skips the load time entirely.
Why speech is detected before transcribing (VAD)
Before feeding the audio to the model, the tool automatically detects the segments where there's actually speech and discards silence. This isn't just a speed optimization: Whisper tends to "hallucinate" repeated text over long stretches of silence or constant background noise, generating loops of the same phrase over and over. Filtering out silence beforehand prevents that on long recordings with pauses.
Why picking the language beats leaving it on automatic
Whisper's automatic language detector is less accurate on short clips, especially with the lighter models (tiny and base). If you know what language the audio is in, selecting it manually noticeably improves transcription accuracy compared to leaving it on "auto-detect". The picker defaults to your own session's language, and automatic mode stays available for anyone who needs it.
The real limits of AI transcription
Like any speech recognition model, Whisper can make mistakes with strong accents, multiple people talking at once, or heavy background noise. It isn't infallible — always review the result before treating it as final, especially for proper names or numbers. If the result looks off, two adjustments usually help: try the max-quality model, or force the language instead of leaving it on automatic.
Frequently asked questions
Is there a minute limit? No, the real limit is your device's memory and power, not an artificial cap set by the tool.
Is the audio uploaded to a server? No. Decoding, voice detection and transcription all happen entirely in your browser.
Does it work with strong regional accents? With varying results — Whisper is robust for common accents but can struggle more with very strong or underrepresented accents in its training. The max-quality model helps in these cases.
Can I transcribe video? Yes, the tool accepts MP4 and M4A files in addition to regular audio formats, extracting the audio track automatically.
Transcribe any audio to text for free with audio to text, running Whisper locally in your browser, with no minute limits and nothing uploaded.