Speech to Text

What it does

This speech to text tool turns an audio or video file into a written transcript, or into SRT or VTT subtitles, without uploading anything. It runs OpenAI’s Whisper speech recognition model directly in your browser, so you can transcribe interviews, meetings, lectures, voice memos and podcasts privately and for free. Whisper can also translate speech in other languages straight into English text.

Pick a file and the page decodes its audio track with the browser’s Web Audio API, resamples it to 16 kHz mono (the format Whisper expects), and transcribes it in a background worker in 30-second windows with a short overlap, so long recordings work too. The result appears in the output box, ready to copy or download as .txt, .srt or .vtt.

How to use

1. Choose a model: Tiny is the fastest, Small is the most accurate, and Base (the default) is a good balance.
2. Press Download model and start. The first time, the model downloads with a progress bar; after that your browser loads it from its cache in a moment.
3. Drop an audio or video file onto the page or press Choose file. MP3, WAV, M4A, AAC, OGG, Opus, FLAC, WebM and MP4 work in most browsers, up to 500 MB. The file name, duration and size are shown.
4. Transcription starts automatically and shows how many segments are done. Press Transcribe again after changing the language or task.
5. Switch the output format between plain text, SRT and VTT at any time; the transcript does not need to be redone. Then Copy or Download it.

The spoken language is detected automatically from the first 30 seconds. Setting it yourself is faster and avoids mistakes with short clips. Choose Translate to English to get English text from speech in another language. The model, language, task and format are saved in the page address, so the Share button links to the same settings; your file is never part of the link.

Example

A two-sentence clip exported as SRT looks like this:

1
00:00:00,000 --> 00:00:03,240
Welcome back to the show.

2
00:00:03,240 --> 00:00:07,800
Today we are talking about speech recognition.

VTT uses the same cues with a WEBVTT header and a dot before the milliseconds, which is the format HTML5 video players read.

Privacy, accuracy and requirements

Everything after the model download happens on your device, which makes the tool suitable for confidential recordings. It needs a recent Chrome, Edge, Firefox or Safari with WebAssembly. Where WebGPU is available (Chrome and Edge on most desktops) it is used automatically and is much faster; otherwise the tool falls back to WebAssembly. Long files need memory: an hour of audio uses a few hundred megabytes once decoded, and the Small model needs more than Tiny or Base, so close other heavy tabs on phones. Keep the tab open until it finishes.

Whisper is good with clear speech and handles accents and background noise reasonably well, but it can mishear names and technical terms, and on silence or music it sometimes repeats or invents a phrase. Review the transcript before you publish it. The tool does not identify speakers, and video files are transcribed only if your browser can decode their audio codec.