How to use it
Give it some speech
Choose an audio or video file up to 30 minutes long, record 15 seconds with the microphone, or share a browser tab that is playing the audio. Sharing a tab works in Chrome and Edge on a computer. Ten seconds or more of one person talking works well.
Wait for the first download
The first time you detect something, your browser downloads the language model and the code that runs it, about 85 MB in all, and keeps them. Later visits skip the download and start in a few seconds.
Read the answer
You get the most likely language with a score, and the next four guesses. A low score means the clip was hard: music, shouting, several voices or very little speech.
What happens to your audio
Nothing leaves the tab. The file is opened by your own browser, turned into sound samples in memory and handed to a model that runs on your computer. There is no upload and no account, and closing the tab discards everything.
You can check this yourself. Detect one clip so the model downloads, switch Wi-Fi off, keep the page open and detect another. It still works. Opening the page itself needs a connection. Apart from the page, the only downloads are the code library, from the jsDelivr network, and the model files, from Hugging Face. Those two services see your IP address and browser, as any website does, and nothing about your audio. Like the rest of this site, the page counts visits and finished detections without cookies, and those counts never include the audio, the file name or the result.
What it can and cannot tell you
- It knows 99 languages, from Afrikaans to Yoruba, and listens to up to three 30 second stretches of a longer clip.
- It names the language, not the dialect or the country. Brazilian and European Portuguese both come back as Portuguese.
- Close relatives get mixed up: Norwegian and Danish, Czech and Slovak, Serbian, Croatian and Bosnian, Malay and Indonesian. Check the second guess.
- It needs speech. Songs, background music and crowd noise pull the score down.
- When two languages are spoken in one clip, you get whichever one dominates.
Once you know the language
Naming the language is usually the first step. The next one is understanding what is being said. Fraze is our desktop app for that: it listens to whatever your computer plays, writes captions as people talk and shows a translation under each line. Like this page, it does the work on your computer, so the audio is not uploaded.
It is a paid app after a free first 30 minutes: $29.99 once or $2.99 a month. It captions 25 European languages, more on macOS 26 or later, and translates about 50 languages to or from English. If the language you just found is outside that list, the result above says so.
Questions
Is my audio uploaded anywhere?
No. The language model runs inside your browser, and the audio is only ever held in this tab's memory. Once the model has downloaded you can turn your connection off, keep the page open, and the detector still works.
How do I find out what language a video is in?
In Chrome or Edge on a computer, play the video in a browser tab, press Listen to a browser tab, pick that tab and tick the box to share its audio. The page listens for 15 seconds and names the language. In other browsers, play it out loud and use the microphone. For a video file on your computer, choose the file instead.
How accurate is it?
On ten seconds or more of clear speech it is usually right. It gets less sure with short clips, music, noise and closely related languages, which is why it shows a score and four runners-up rather than one bare answer.
Which languages can it detect?
99, covering most widely spoken languages and many smaller ones, including Arabic, Bengali, Chinese, Dutch, Farsi, Greek, Hebrew, Hindi, Indonesian, Japanese, Korean, Polish, Swahili, Tamil, Thai, Turkish, Ukrainian, Urdu and Vietnamese.
Why does it download about 85 MB the first time?
That is the language model, about 80 MB, and the code that runs it. Because it runs on your computer and not on a server, your browser has to fetch it once. It is stored in the browser's cache, so later visits skip the download.
Which model does it use?
OpenAI's open source Whisper model, the base size in 8-bit form, run with the transformers.js library on WebAssembly. The model files come from Hugging Face and the library from jsDelivr. Whisper is published under the MIT license.