Filery

Speech to Text Beta

Transcribe recordings with Whisper. Timestamps and captions included.

Worth knowing before you startThe Whisper model is 75–150 MB depending on the size you choose. On pure WebAssembly it runs at roughly real-time (one minute of audio takes roughly one minute to transcribe); WebGPU acceleration is much faster where available. Long recordings should be chunked into sections of 30 minutes or less for best results.
Choose audio or video file or drop it anywhere on this page MP3 · WAV · OGG · OGA · M4A · AAC · FLAC · OPUS · WMA · AIFF · +5 Processed on your device — never uploaded

How to speech to text

  1. Add an audio or video file — MP3, WAV, M4A, OGG, MP4, MOV and others are all accepted.
  2. Choose the model size. Base is the best starting point. The model downloads once and is cached.
  3. Select the output format — SRT or VTT if you need subtitles, TXT for a plain transcript.

About this tool

Whisper is OpenAI's open-source transcription model, released in 2022, and it shifted the field. Before it, accurate automatic transcription required either an expensive commercial API or a dedicated GPU. Whisper is small enough to run in a browser via WebAssembly, and the quality on clear speech is good enough to replace most manual transcription workflows.

The model comes in several sizes. Tiny (75 MB, roughly real-time on WASM) is sufficient for clear recordings in English. Base and Small are meaningfully more accurate and worth the extra seconds of first-load time for anything that matters. WebGPU acceleration — available in Chrome and Edge on most modern hardware — can reduce processing time to a fraction of real-time.

Because everything runs locally, the tool is appropriate for sensitive recordings: therapy sessions, legal depositions, confidential interviews, medical notes. The audio never leaves your machine. If you want to verify this, open your browser's Network tab before loading a file — you will see the model download once, and nothing else.

Questions

How accurate is Whisper?
On clear speech in a supported language, Whisper is competitive with commercial transcription APIs. Accuracy drops for strong accents, overlapping speakers, heavy background noise and technical jargon. The Small model is noticeably better than Tiny on all of these, at the cost of a longer first-load.
Can it handle multiple speakers?
Whisper does not perform true speaker diarisation — it transcribes what was said but does not reliably label who said it. The JSON output includes word-level timestamps, which can be fed into separate diarisation tools if speaker attribution matters to you.
Is there a recording length limit?
No hard limit, but the practical advice is to chunk recordings longer than 30 minutes. Very long files take a long time on WebAssembly, and the model's context window means accuracy can degrade in the middle of very long continuous segments. Split at natural breaks — interview questions, chapters, scene changes.
Will my private recordings be uploaded?
No. Transcription happens entirely in your browser. The audio does not leave your device. This matters for medical consultations, legal proceedings, confidential meetings, and anything else you record that you would not want on a third-party server.