Speech to Text
Transcribe audio or video to text and subtitles, on your device. Runs in your browser.
About this speech to text
This speech to text tool turns an audio or video file, or a recording you make on the spot, into text with timestamps. It is for anyone with a lecture, interview, voice note or video that needs typing up or subtitling, and it runs entirely in your browser.
It uses Whisper, the open speech recognition model from OpenAI, in its base size. The model downloads from this site once (about 94 MB, or 176 MB for the faster graphics-chip version), your browser keeps it, and from then on every file is transcribed on your own device. No upload, no account, no daily minute allowance.
You get plain text, SRT or WebVTT subtitles, and a timestamp next to every line that jumps the player to that moment. Whisper can also translate what it hears into English.
Features
Files and live recordings
Drop in MP3, WAV, M4A, OGG, FLAC, MP4, MOV, MKV or WebM, or press Record and speak into your microphone. A level meter and timer show that it is listening.
Long audio in 30-second steps
Whisper hears 30 seconds at a time. The tool walks through your file window by window, re-reads any sentence a window cut in half, and shows text as it goes, with time left and Cancel.
Automatic language detection
On Auto-detect, Whisper scores the first 30 seconds with sound against its 99 languages and shows its best guess with a probability. Or pick one of 43 languages yourself.
Translate to English
Switch the output to Translate to English and Whisper writes English text for Spanish, Hindi, Bengali or any language it knows, with the same timestamps.
TXT, SRT and VTT export
Copy the text, or download it as TXT (with optional [m:ss] stamps), SRT for most video editors and YouTube, or VTT for web players. Overlapping times are fixed so the files load cleanly.
Graphics chip when available
With WebGPU, Whisper runs on your graphics chip, about five times faster than on the processor in our tests. If the GPU fails, the tool switches to the processor by itself.
How to use the speech to text
- 1Choose an audio or video file, or record from your microphone.
- 2Pick the language or leave Auto-detect, and choose same language or English.
- 3Press Transcribe, then copy the text or download TXT, SRT or VTT.
When to use it
Lecture and class recordings
Turn a 50-minute lecture into notes you can search. If you only need one part, cut that section out first so the run is shorter.
Interviews you can't upload
When you promised a source the recording stays private, a tool that never uploads it is the easy choice.
Subtitles for a video
Drop the MP4, transcribe, and download the SRT. For a big file you can also pull the audio track out of the video first; the result is the same and the file is smaller.
Voice notes and memos
Record a thought, or use the voice recorder to capture a longer memo, then turn it into text you can paste into an email.
How accurate is Whisper base?
Accuracy depends far more on the recording than on the tool. Clear English from one speaker close to a microphone usually comes back nearly word for word, with punctuation and capital letters. Phone recordings in a café, two people talking over each other, or music under the voice all add mistakes.
The base model is the second smallest of Whisper's five original sizes, with 74 million parameters. It fits in a browser, but it is less accurate than the large models cloud services run. That is the price of keeping your audio local.
Expect most errors on names, brand names, numbers and technical terms. A quick pass with find and replace fixes a name that is misspelled the same way throughout.
Does it work for Bengali and other languages?
Yes, but with a clear warning for Bengali. Whisper was trained on far less Bengali audio than English, and the small models show it: the text is often understandable but has many wrong words, odd spellings and missed endings. Treat a Bengali transcript as a rough draft to correct, not a finished document.
Languages with more training data, such as Spanish, French, German and Japanese, do noticeably better. If auto-detect guesses wrong on a short or noisy clip, pick the language yourself.
Why not use the browser's own speech recognition?
Chrome has a built-in speech recognition API, and many simple dictation sites use it. In Chrome that API streams your microphone audio to Google's servers to be recognized, so it is not private, it needs a connection, and it cannot read files at all.
This tool skips it on purpose. Whisper runs inside the page, so nothing is sent anywhere.
SRT vs VTT vs TXT
All three hold the same words. Pick by where they go next.
- SRT: numbered blocks with times like 00:01:02,500. Accepted by YouTube, Premiere Pro, DaVinci Resolve, CapCut and VLC.
- VTT (WebVTT): the W3C caption format for the HTML video element. Times use a dot, 00:01:02.500, and the file starts with WEBVTT.
- TXT: plain text for notes and documents. Turn on timestamps to get [1:02] before each line, or leave them off for paragraphs that break at pauses longer than two seconds.
Why use Mavros
| Mavros | Typical upload sites | |
|---|---|---|
| Where the audio goes | Stays on your device | Uploaded to the site's servers |
| Free limit | No minute or file count limit | 10 to 30 minutes a day, or 3 files |
| Exports | TXT, SRT, VTT, copy | TXT free, subtitles often paid |
| Translate to English | Yes | Paid add-on or missing |
| Model size | Whisper base (smaller, less accurate) | Often larger cloud models |
Tips
- Record close to the microphone in a quiet room. Distance and echo cost more accuracy than a cheap mic does.
- If the speakers use one language, choose it instead of Auto-detect. It saves a step and avoids a wrong guess on a noisy intro.
- Press Download the model now while you find your file, so the first-run wait is over by the time you need it.
- Made a voice-over with our text to speech tool? Run it back through here to catch words the voice got wrong.
- Plan your time. With WebGPU, an hour of audio takes roughly half an hour; on the processor alone it can take two hours or more. Keep the tab open; switching to another tab is fine.
- Whisper sometimes invents a line such as "Thank you." over silence or music. Silent stretches are skipped and repeated lines are collapsed, but give the end of the transcript a quick look.
Questions
Is this speech to text tool really free?
Yes. There is no sign-up, no daily limit and no paid tier. The model runs on your own device, so there is no server cost per minute to pass on to you.
Is my audio uploaded anywhere?
No. The page downloads the Whisper model from this site once, then decodes and transcribes your file inside the browser tab. Your recording is never sent to a server, and closing the tab discards it.
How long does transcription take?
With WebGPU in Chrome or Edge, expect roughly 25 to 30 seconds of work per minute of speech, even on built-in laptop graphics. On the processor alone it is slower, around two to three minutes per minute. Long silences are skipped, so they cost almost nothing.
Can it transcribe Bengali (Bangla)?
Yes, but the result needs checking. Whisper base was trained on little Bengali audio, so expect wrong words, especially names and numbers. Pick Bengali in the language list rather than relying on auto-detect for short clips.
Can I make subtitles for YouTube?
Yes. Transcribe the video file and download the SRT. YouTube Studio accepts SRT under Subtitles, Upload file, With timing. VTT works too and suits web players.
What file types can I use?
Most audio and video files: MP3, WAV, M4A, AAC, OGG, Opus, FLAC, MP4, MOV, MKV and WebM. It uses your browser's own decoders, so a codec your browser can't play, such as some WMA files, needs converting first.
Does it tell speakers apart?
No. Whisper writes what is said but not who said it. For interviews, add the names while you read through the transcript.