This speech to text tool transcribes recordings and videos into text and ready-made subtitles using Whisper, the open speech recognition model from OpenAI, running entirely in your browser. Download a model once — Tiny for speed, Base for most recordings or Small for the best accuracy — straight from Hugging Face into your browser, then drop any number of MP3, WAV, M4A or MP4 files and they are transcribed one after another on your own device. Each result downloads as a plain transcript with paragraphs, an SRT subtitle file for video editors and YouTube, or a WebVTT file for the web, and a whole batch downloads as one ZIP. It detects the language automatically, supports Urdu, Hindi, Arabic and dozens more, and can translate speech into English. Your recordings are never uploaded, and you can delete the model from your browser at any time.
| Model | Approx. download | Best for |
|---|---|---|
| Tiny | ≈ 45 MB | Quick drafts of clear English speech |
| Base | ≈ 80 MB | Most interviews, lectures and videos |
| Small | ≈ 250 MB | Accents, noisy audio and other languages |
No. The Whisper model runs in your browser, so recordings and transcripts stay on your device. Only the model itself is downloaded once, from Hugging Face.
Base suits most clear recordings. Tiny is fastest for short, clear English. Small is the most accurate, especially for accents and other languages, but is slow without WebGPU.
Drop the video file, transcribe it and download the SRT file. Upload it to YouTube or import it into your video editor alongside the video.
Use Clear downloaded models. You will be asked to confirm, and you will need to download the model again next time.
Usable but less accurate than English, especially on Tiny and Base. Choose Small and check the transcript.
Transcription uses OpenAI's Whisper (MIT licence) through Transformers.js in a Web Worker, on WebGPU when available and WebAssembly otherwise. The visitor downloads the chosen model (Tiny, Base or Small) once, directly from Hugging Face into the browser's Cache Storage; it can be removed at any time from the Downloaded models panel. Audio and video files are decoded by the browser straight to 16 kHz mono. Long recordings are cut into roughly five-minute segments at the quietest 50 ms near each boundary so words are not split, and each segment runs through Whisper with 30-second windows, a 5-second stride and segment timestamps. Timestamps are shifted back to the recording's timeline, long cues are split by text length and wrapped to two lines of about 42 characters, and the result is written as plain text with paragraphs at pauses, SRT or WebVTT. Whisper can also translate any supported language into English.