About Live Subtitle Translator
Live Subtitle Translator turns the speech in any Chrome tab into subtitles as it plays: a video, a call, a lecture or a stream. The original line appears within about a second (Japanese about two and a half), and the translation lands under it in large type when the sentence ends. It hears 28 spoken languages and detects which one is talking. Words the model is unsure of are marked, nothing is guessed in silence, and every line goes into a transcript you can save as SRT, VTT or text.
What it does
- Live subtitles for any Chrome tab: videos, calls, lectures and streams
- Translation into your language when each sentence ends, in large type under the original line
- 28 spoken languages from one speech model, detected automatically, with an offer to lock the language once it is heard
- Honest output: words not yet final are shown muted, unsure words are underlined, and nothing appears during silence
- A transcript with times on the capture clock, searchable, saved as SRT, VTT or plain text with the original, the translation or both
- Float the subtitles in a small window that stays over other windows
- Works offline after the speech model's one-time download
How to use Live Subtitle Translator
Open the tool in desktop Chrome
Press Share a tab, or Try a sample to watch a Korean clip get subtitled first. The first use offers the speech model, a one-time download with its size shown before anything starts.
Pick the tab with the video or call
In Chrome's share window, choose the tab and keep Also share tab audio on. The tab keeps playing as usual.
Read the subtitles
The original line appears as it is spoken and the translation follows when the sentence ends. Pick the language you want under Subtitles in; lock the spoken language when the tool offers it.
Save the transcript
Press Export and save every line as SRT or VTT subtitles or as plain text with times.
Frequently asked questions
Which browsers does it work in?
Chrome on a computer (Windows, Mac, Linux or a Chromebook), and other desktop browsers built on Chromium. Firefox, Safari and phones can't share another tab's sound with a page, so the tool doesn't run there and nothing downloads.
Which languages can it hear?
28 languages (32 counting regional variants such as US and UK English): English, Spanish, French, German, Italian, Portuguese, Dutch, Japanese, Korean, Russian, Ukrainian, Turkish, Arabic, Hindi, Vietnamese, Polish, Swedish, Czech, Norwegian, Danish, Bulgarian, Finnish, Croatian, Slovak, Hungarian, Romanian, Estonian and Mandarin Chinese. Mandarin gets its own speech model, offered the first time Mandarin is heard. Hindi is marked rough in the language list, and Mandarin too until its own model is on this computer. Translation covers the languages your browser's on-device translator offers.
How big is the download, and does it work offline?
The speech model is a one-time download of about 790 MB, shown before you agree to it, plus a 24 MB runtime that Loft's other AI tools share. Mandarin has its own model of about 170 MB, offered only when Mandarin is first heard. After that, subtitles work with no connection; translation runs on your computer too.
How accurate is it?
Measured on clear read speech (Google's FLEURS recordings), words wrong in 100: Spanish about 6, Portuguese 10, English 10 when you pick English (12 when it detects the language), French 11, German 12, Vietnamese 14, Russian 15, Arabic 16, and Hindi 24, which is why Hindi is marked rough. Characters wrong in 100: Korean 9, Japanese 17, and Mandarin 13 with the Mandarin model, 28 without. Other languages are not measured yet. Real videos with music, several people talking at once or a noisy room do worse, and loud background chatter can add a stray word. Unsure words are underlined instead of hidden.
What does it need from my computer?
A graphics chip Chrome can use (most laptops from the last few years have one) and about 820 MB of browser storage, or about 1 GB with the Mandarin model. If Chrome only has a software stand-in for graphics, the tool says so before any download.
Why is Japanese a little slower?
The speech model waits for more context before it writes Japanese, so its first words show about two and a half seconds after speech starts, against about a second for English.