Kit-Bin
Donate

Transcribe Audio to Text

Get a text transcript and an SRT subtitle file. Processed entirely on your device.

Processed entirely in your browser. Never uploaded to a server.

Drop an audio file here or click to browse

Processed entirely in your browser. Never uploaded to a server.

Audio only for now (MP3, WAV, M4A, OGG, FLAC). Not video files.

This tool turns spoken audio into a written transcript, plus a timed SRT subtitle file you can drop straight into a video editor. Drop your audio file above, wait for the model to load and the audio to process, then copy the transcript or download it as a plain-text file or an SRT.

Every commercial transcription service you've probably seen charges per minute of audio specifically because it runs the actual speech recognition on their servers — your file has to leave your device for them to process it and bill you for it. This tool runs the same kind of speech recognition model (OpenAI's Whisper) directly in your browser instead, so there's no per-minute meter and no upload step: the audio is decoded and transcribed on your own device, in this tab, and is never sent anywhere. The only network activity involved is a one-time download of the recognition model itself (around 150MB, cached by your browser after the first run) — that's the model's weights, not your file.

FAQ

How long does transcription take?
It depends on your device and audio length. On a machine with WebGPU (most recent Chrome/Edge on a discrete or capable integrated GPU), a few minutes of audio typically finishes in well under the audio's own length. Without WebGPU the model falls back to a CPU-only WASM path, which can take noticeably longer than the audio itself — a 10-minute recording taking several minutes is normal on an older laptop or phone, not a sign something is broken.
What languages does this support?
Whisper base is a multilingual model and can transcribe dozens of languages, but this tool is tuned and tested for English. Other languages will often produce a usable transcript, but accuracy is not something we've verified beyond English — treat non-English results as a rough draft rather than a finished transcript.
Is my audio uploaded anywhere?
No. The audio file is decoded and transcribed entirely in this browser tab. Separately, the speech recognition model itself (around 150MB) downloads once from Hugging Face's CDN the first time you use this tool, and your browser caches it after that — that one-time model download is not the same thing as your audio file, which never leaves your device.
How accurate is it with background noise, multiple speakers, or accents?
Clear single-speaker audio with minimal background noise gets the best results. Accuracy drops with heavy background noise, overlapping speakers talking at once, and strong or unfamiliar accents — this model doesn't separate or label different speakers, so a multi-person conversation comes back as one continuous transcript, not one broken out by who said what. Treat the output as a strong first draft, not a certified transcript.
Can I transcribe a video file?
Not directly on this tool — v1 only accepts audio files (MP3, WAV, M4A, OGG, FLAC, and similar). Decoding a video file's audio track reliably in-browser across every browser (Safari in particular) isn't solid enough yet to promise here. Pull the audio out first with Extract Audio from Video, then drop the result into this tool.

Related tools: MP3 to WAV, Extract Audio from Video, Trim Audio.