How to Transcribe Audio Without Uploading It
On-device speech-to-text means your recording never leaves your computer. Here's how it works, and which model to pick for what you're transcribing.
Published July 28, 2026
Why on-device matters
Most transcription tools work by uploading your audio to a server, running it through a cloud speech model, and sending the text back. That's fine for a lot of use cases, and it's a real problem for others — an interview under NDA, a legal or medical conversation, an HR meeting, anything where the recording itself is sensitive. On-device transcription sidesteps the question entirely: there's no server in the loop to trust.
How it works
nah.tools' transcription tool runs OpenAI's Whisper model converted to ONNX format, executed in your browser using WebGPU when available, with a WebAssembly fallback. Model assets download from Hugging Face, and the page and runtime need assets too. Browser caching can reduce later downloads but does not guarantee offline access. Audio processing runs locally; the recording is not sent to a transcription service.
Which model to pick
Download size varies with the selected model and execution backend. Allow time and storage for model and runtime assets; the download indicator reports progress.
- English Fast — English-only, the smallest model and a fast starting point. Good default for a quick English voice memo.
- Multilingual — supports non-English languages; accuracy still depends on language and recording quality.
- High Accuracy — a larger model to try when the smaller models make errors. It does not identify speakers, and its output still needs review.
Steps
- Open /audio/transcribe and pick a model using the options above.
- Drop in your audio or video file — model and runtime assets download when needed.
- Wait for processing; larger models and longer recordings take more time, especially on lower-powered devices.
- Export as plain text, SRT/VTT subtitles, or a formatted PDF of notes.
If the recording is noisy or has a wide dynamic range, running it through audio normalization first may help with uneven volume. Normalization does not remove background noise, and it does not guarantee better recognition. Always review the transcript against the recording.
Frequently asked questions
Does any of my audio get uploaded?
No. The page, runtime, and model assets may download over the network; browser caching can reduce later downloads. Your recording is processed locally and never leaves the device.
Which model should I pick?
Start with English Fast for English speech, use Multilingual for other supported languages, and try High Accuracy if smaller models make errors. Larger models need more resources; none guarantees an accurate transcript or identifies speakers.
Do I need an internet connection?
Plan to be online to load the page, runtime, and model assets. Processing is local once loaded, but cached files can be evicted and offline startup is not guaranteed. Test offline use in advance if you need it.
What formats can I export?
Plain text, SRT and VTT subtitle files, and a formatted PDF of notes.