Audio/Video Transcriber
Transcribe audio or video to text and generate SRT/VTT subtitles with on-device Whisper AI. Free, private, no upload. Runs on your device — nothing is uploaded, nothing is stored.
Drop an audio or video file here
Supports MP3, WAV, M4A, OGG, MP4, and WebM. One file at a time.
Runs OpenAI Whisper entirely in your browser via WebGPU (with a WASM fallback). The model downloads once from a CDN on first use, then transcribes on-device — your audio is never uploaded.
Audio/Video Transcriber turns a recording into text on your own machine. Drop in an MP3, WAV, M4A, OGG, MP4 or WebM file and it produces a full transcript plus timestamped segments you can export as SRT or VTT subtitles, or as plain text.
The speech recognition is Whisper — specifically the base checkpoint — running through Transformers.js in the browser. The first time you use it, the model files download and are cached; after that they're reused. The audio itself is decoded to 16 kHz mono in the page and fed to the model in thirty-second chunks with a five-second overlap so words that fall on a chunk boundary aren't lost.
No API key, no per-minute charge, no upload of the recording.
Why this runs on your device
This is the tool in the catalogue where on-device processing genuinely changes what you're allowed to do. The recordings people want transcribed are therapy sessions, medical dictation, HR investigations, legal depositions, journalists' interviews with sources who were promised confidentiality, and internal all-hands calls. Sending any of those to a hosted transcription API means a copy on someone else's infrastructure, governed by a retention policy and terms you probably haven't read — and in several professions, that alone is a compliance problem regardless of what the vendor promises. Here the audio never leaves the tab. The model travels to your recording instead of your recording travelling to the model.
- Drop in an audio or video file.
- Press Transcribe — the first run downloads the Whisper model, with a progress bar.
- Wait for the transcript and its timestamped segments to appear.
- Copy the text, or download it as TXT, SRT or VTT.
- It runs Whisper base, the small end of the family. A hosted API is almost certainly running something far larger, and the difference shows: expect more errors on strong accents, background noise, crosstalk, proper nouns and technical vocabulary. There are no speaker labels — overlapping speakers come out as one undifferentiated stream of text.
- The first run downloads the model from a CDN. Your audio isn't uploaded, but the tool isn't usable offline until those files are cached, and on a slow connection the wait before transcription even starts is real.
- Speed depends entirely on your hardware. On a browser with WebGPU it's reasonably quick; falling back to WebAssembly can be slower than real time, meaning an hour of audio can take more than an hour. There's no progress indicator during transcription itself and no way to cancel — only the model download shows progress.
- The whole file is decoded into memory before transcription starts, so long recordings can exhaust the tab on a modest machine. There's also no language selector and no translate mode — Whisper detects the language itself and you can't override it — and the segment timestamps are approximate, so subtitles usually need a nudge in a subtitle editor before use.