Audio to Text: Transcribe Recordings and Make Subtitles
Transcribe an audio or video file with the Whisper speech recognition model, right in your browser, and save the text or SRT and VTT subtitles. The recording is not uploaded.
Drop an audio or video file here, or choose one
MP3, M4A, WAV, OGG, OPUS, FLAC, WebM and MP4 work in most browsers. Other formats work if your browser can play them.
This recording is longer than an hour. Transcribing it can take as long as the recording or longer, and the browser may run out of memory with very long files. If that happens, split the file into parts.
Your browser reports mobile data or data saver mode. Waiting for Wi-Fi is recommended.
Choose the language and model, then add a file.
Notes
- The recording is decoded and transcribed in your browser. It is not uploaded to LeedyBox or anywhere else, and the tool can only contact leedybox.com while it works.
- The first time, your browser downloads the speech recognition engine and the model you chose from leedybox.com: about 90 MB for base, 250 MB for small or 750 MB for large-v3-turbo. They are then saved in this browser's storage, so later visits do not download them again. "Delete saved models" under the model choice removes them; clearing the site's data in the browser does too.
- Base and small run on your processor, large-v3-turbo on your graphics processor, one window of 30 seconds after another. On our test computer, base took about half the length of the recording, small about one and a half times it and large-v3-turbo about a third; phones and older computers are slower. Keep this tab open until it finishes.
- Clear speech from one speaker close to the microphone works best. Background music, noise, people talking at the same time, accents and specialised words lead to mistakes, and names are often misspelled. Whisper can also repeat a phrase or write words that were not said, mostly in silent or noisy parts.
- One language is used for the whole file. Passages in another language are written in that language as well as the model can, or translated. Speakers are not told apart.
- Subtitles follow the segments Whisper finds, usually a sentence or a phrase each. Check the timing in a player before you publish them; long segments may need to be split for reading.
- Very long recordings need a lot of memory: an hour of audio takes about 220 MB once decoded, plus the model. If the browser stops the page, split the file into shorter parts.
Something wrong? If a file cannot be opened, the model does not load, the page stops or the transcript is much worse than expected, please email us. Tell us the model and language you chose, the file type and length, and your device and browser. Do not send the recording.
Does a sentence on this page sound unnatural in your language? Let us know that too.
How to use
- Choose the language spoken in the recording, or leave "Detect automatically".
- Choose the model: base is the fastest, small is more accurate, and large-v3-turbo, offered when your browser supports WebGPU, is the most accurate.
- Drop an audio or video file on the box or choose it, then press "Start transcribing". The first time, the engine and the model are downloaded from leedybox.com.
- Watch the text appear as each part of the recording is transcribed. You can stop at any time and keep the text so far.
- Check the transcript against the recording, then copy it or download it as a .txt file, or as .srt or .vtt subtitles for a video.
Frequently asked questions
Is my recording uploaded to a server?
No. The file is decoded and transcribed in your browser. LeedyBox only sends the speech recognition engine and the model to your browser; the audio and the transcript are not sent anywhere, and the part of the page that runs the model is blocked from contacting any other site.
Which languages does it support?
The 99 languages of the multilingual Whisper models, including English, Korean, Spanish, Japanese, Chinese, German, French, Portuguese, Russian, Vietnamese, Indonesian, Italian, Thai and Turkish. Accuracy varies a lot between languages: it is best for widely spoken languages with plenty of training data.
Which model should I choose?
On a computer whose browser offers large-v3-turbo, use it: in our tests it made no mistakes where base and small did, and it was the fastest because it runs on the graphics processor; it downloads about 750 MB once. Otherwise start with base (about 90 MB, about half the length of the recording on a recent computer) and switch to small (about 250 MB, about three times slower) when the transcript has too many mistakes. On phones, use base.
How long can the recording be?
There is no fixed limit, but processing time grows with the length and everything happens on your device. Recordings of up to an hour work well on a computer. Longer files can take hours and may run out of memory; splitting them into parts is safer.
What is the difference between SRT and VTT?
Both are subtitle files with the text and the time it appears. SRT (SubRip) is the most widely accepted format for video editors and players. VTT (WebVTT) is the W3C format for subtitles in web pages and HTML5 video. Choose the one your player or platform asks for.
Why are some words wrong or invented?
Whisper guesses the most likely text from the sound. Noise, music, overlapping voices, accents, names and technical terms cause mistakes, and in silent or very noisy parts it sometimes repeats phrases or produces text that was not said. Choosing the language, using the small model and recording closer to the microphone help.
Why is large-v3-turbo not offered in my browser?
It runs on the graphics processor through WebGPU, a browser feature for computing on the GPU. Recent versions of Chrome and Edge on Windows, macOS and ChromeOS, and Safari on recent Macs, offer it; some other browsers and many phones do not. On a processor alone this model would take many times the length of the recording, so it is shown only where WebGPU is available.