Audio File Transcriber
Upload a pre-recorded audio file and get a text transcript, using on-device speech recognition.
What is File Transcriber?
A transcription tool for pre-recorded audio and video: interviews, voice notes, lectures and meeting recordings, converted to text (and subtitles) without uploading the file.
How it works
The file is decoded locally, resampled to 16 kHz mono, and transcribed by Whisper running in WebAssembly on your device in 30-second chunks with a 5-second overlap. Timestamps from the model become SRT/VTT subtitle tracks.
- Load your file. Drag a file onto the drop zone or pick one from your device. It is read locally by the browser and never uploaded.
- Set the options. Adjust the handles, format, quality or threshold controls — the preview updates as you change them.
- Process and download. Run the tool, then download the result. Compare the reported before/after size for compressors and converters.
Examples
A 42 MB podcast episode → 9 MB MP3
Trim the intro, normalise to −16 LUFS, and export at 96 kbps mono — a typical interview lands around a fifth of the original size.
A phone video with the wrong orientation
Rotate 90° and the width/height swap in the output, so it plays upright in every player instead of relying on metadata flags.
Common mistakes
Expecting instant encoders
Video encoding runs in single-threaded WebAssembly, so a clip takes roughly its own length (or a little longer) to re-encode. Stream-copy operations are near-instant by comparison.
Re-encoding when a copy would do
Trimming, removing audio and joining same-codec clips can copy streams without re-encoding — pick the copy option when it is offered to avoid any quality loss.
Ignoring the first-load download
The media engine is about 31 MB and is fetched on first use, then cached. The first conversion on a connection is slower than every later one.
Frequently asked questions
How accurate is it?
Whisper tiny.en is strong on clear speech in quiet conditions and struggles with heavy accents, overlapping speakers or background music. Larger models are more accurate but far slower in a browser.
Which languages are supported?
This build uses the English-only tiny.en model. English audio works best; other languages will produce poor results.
How long does transcription take?
Roughly real time on a modern laptop for a first run, plus a one-time 39 MB model download that is cached afterwards. Nothing is sent to a server.
How is this different from the live Speech-to-Text tool?
That tool listens to your microphone in real time via the browser's speech recognition. This one accepts an existing file, works offline and can produce timestamped subtitle files.