How to convert MP3 to text
- Open the MP3 file.
- Choose a model: Base suits most recordings; Tiny is fastest for English; Small is the most accurate.
- Leave Language on automatic detection, or pick the spoken language.
- Select Transcribe, then copy the text or download TXT, SRT or VTT.
Worked example
A 10-minute MP3 is split into sections of 20 to 30 seconds, cut at quiet moments, and the text appears section by section with timestamps; the progress line counts the sections as they finish.
Formats and limits
| Input | MP3, WAV, M4A, Ogg, FLAC, WebM, MP4, MOV |
|---|---|
| Output | TXT, SRT, VTT, JSON |
| Limits | First use downloads 44 MB (Tiny, English), 80 MB (Base) or 252 MB (Small) from Hugging Face, more on WebGPU; cached afterwards. Up to about 2 hours, processed in sections of up to 30 seconds. |
| Engine | Whisper (MIT) via transformers.js in a Web Worker: WebGPU when available, otherwise WebAssembly; cuts at quiet points, silent sections skipped |
Limitations
- Accuracy depends on the recording; background noise, music and overlapping speech lower it.
- Up to about 2 hours of audio.
- The first run downloads the model from Hugging Face (44 to 252 MB).
Questions
Is the MP3 uploaded for transcription?
No. The audio is decoded and transcribed in your browser. Only the model files are downloaded, once, from Hugging Face.
How big is the first download?
44 MB for Tiny (English only), 80 MB for Base or 252 MB for Small, and more when running on WebGPU. Later runs use the copy cached in your browser.
Can it label who is speaking?
No. The text comes with timestamps for each segment, but speakers are not identified.
Privacy
On-device model. Processing runs in this browser. The open-source model files are downloaded once from the model host (Hugging Face) and cached; your content is not uploaded.
- Hugging Face: Downloads of the open-source Whisper model files when you ask for them. No audio or text is sent.
This page opens the Audio transcription ready for this task. See the privacy policy for how toolsdocks handles data.