Local speech recognition · No account required
Transcribe audio to text on your device.
Choose a short recording, generate a transcript and review the words before exporting TXT or timed subtitles. Audio stays in your browser.
Up to 50 MB / 5 minutes. Browser-supported WAV, MP3 or MP4/AAC. This tool does not fetch YouTube links, extract embedded subtitle tracks or read on-screen text.
The first run downloads a speech model and runtime from Hugging Face and this site or jsDelivr (tens of MB). Your audio is processed on this device. Processing can take several minutes and uses memory and CPU; keep this tab open.
No audio, transcript or filename is sent to DoneCove or the model host. Model downloads expose normal network metadata to those providers. Your browser may cache model files. Anonymous tool-use counts contain no recording content.
Choose a recording to begin.
Review the generated draft
Speech recognition may omit words, repeat text or invent speech in silence. Timings are approximate. Check the recording before using this result.
For visual timing edits, download SRT and open the subtitle editor. Subtitle editor
From recording to reviewed text
- Choose a browser-supported audio file, up to 50 MB and 5 minutes.
- Choose the spoken language, then start local transcription. The first run downloads the speech model.
- Review the generated SRT, correct words and timing, then download TXT, SRT or VTT.
How accurate is local transcription?
This uses Whisper tiny through Transformers.js. Results vary with language, accent, noise and recording quality. The small model favors a lighter download over maximum accuracy. No claim of human-level accuracy is made.
Can I use a long recording or paste a YouTube URL?
This version handles local clips up to 5 minutes and 50 MB. Longer recordings must be split beforehand. YouTube URL retrieval, OCR, subtitle-track extraction and dubbed audio are not supported.
What if the model cannot download or the browser cannot decode my file?
An error is shown and no completed transcript is claimed. Retry on a stable connection, or try a short WAV file in a desktop browser. Cancel stops the recognition worker.