On-device speech to text
Transcribe runs OpenAI's Whisper model inside your browser through Transformers.js. The audio never leaves your machine: there is no upload step, no server-side queue, and no per-minute transcription bill. The model weights download once and are cached by the browser, so subsequent sessions start quickly even offline.
WebGPU with a WASM fallback
Where the browser exposes WebGPU, inference runs on the GPU and is markedly faster than real time on short clips. Where it does not, the same model runs through WebAssembly on the CPU. The tiny and base checkpoints are offered because they are the sizes that stay responsive in a browser tab; larger Whisper variants are accurate but impractical to load client-side.
Input and output formats
You can drop an audio or video file, or record directly from the microphone. Results export as plain text, SRT and VTT subtitle tracks with segment timestamps, or JSON when you need the raw segment structure for further processing. Whisper's translate mode converts ninety-nine source languages into English.
Accuracy expectations
Small Whisper checkpoints handle clear single-speaker audio well and degrade on crosstalk, heavy accents, background music, and specialised terminology. Timestamps are segment-level, not word-level. For archival or legal work, treat the output as a first draft that a human corrects, which is still far faster than transcribing from nothing.
Frequently asked questions
Is my audio uploaded to a server?
No. The model runs inside your browser and the audio never leaves your device, which is the main reason this tool exists.
Do I need an API key?
No. Transcription is entirely local and costs nothing per minute. Model weights download once and are then cached.
Which formats can I export?
Plain text, SRT and VTT subtitles with segment timestamps, and JSON for programmatic use.
Can it translate?
Yes. Whisper's translate mode converts about ninety-nine source languages into English text.