Turn Hindi speech into text in Devanagari on your own device: drop a recording or a video and get a Hindi transcript, or timed .srt subtitles. Be ready to correct some words. The small Whisper models that run in a browser saw far less Hindi in training than Spanish or German, so the Accurate model is the one to use. Everyday Hindi borrows English freely, and with Hindi selected those words are often written in Devanagari (मीटिंग for meeting), sometimes in Latin letters, not always consistently. Nothing is uploaded.
How it works
- Drop a Hindi audio file or video; Hindi is preselected as the spoken language.
- Choose Accurate (≈77 MB) unless a rough draft is enough; Fast (≈41 MB) makes more mistakes on Hindi. For captions, switch Output to Subtitles before you start.
- Press Transcribe, play back any line that looks wrong, fix it in place, and download .txt, .srt or .vtt in Unicode Devanagari.
Questions
- How accurate is Hindi speech to text here?
- Usable on clear, close-mic speech from one person; weaker on phone calls, crowds, film audio and fast Hinglish. Expect to correct more words than you would in an English transcript. The best result comes from the Accurate model, a quiet recording and Hindi selected, as it already is on this page. If music runs under the voice, separate it first with the vocal isolator.
- Can I get Hindi in English letters (Roman Hindi)?
- No. With Hindi selected the transcript is written in Devanagari and there is no romanised output. Switching the language to English will not produce it either: the model then decodes Hindi audio as English and returns English-looking text, not Hindi in Latin letters. To romanise, transliterate the finished .txt with a separate tool.
- Can it transcribe Urdu?
- Urdu is not one of the 15 languages in the picker. Spoken Urdu and Hindi are close, so the Hindi setting can follow much everyday Urdu speech, but it writes Devanagari rather than Urdu script, and Persian and Arabic loanwords fare worse. For a transcript in Urdu script, this tool is not the right one yet.
- Does my file get uploaded?
- No. The models and all processing run inside your browser on your own device. You can verify this yourself: open your browser’s DevTools, watch the Network tab, and process a file: there are no uploads. After the first visit most tools also work fully offline.
- How accurate is it?
- The models are OpenAI’s Whisper (tiny and base tiers). On clear speech they are very usable; heavy accents, crosstalk and noisy rooms will show errors. The Accurate tier is noticeably better on difficult audio.
- Which languages work?
- Fifteen, picked in the Language setting: English, Spanish, French, German, Italian, Portuguese, Dutch, Russian, Polish, Turkish, Japanese, Korean, Chinese, Arabic and Hindi. The decoder is conditioned on the one you pick, so it transcribes that language rather than approximating it in English — which also means picking the wrong one returns confident nonsense. The weights understand 99 languages; these are the ones we measured as good enough to offer.
- How long does it take?
- Transcription runs on your processor, so speed depends on the machine: expect around a quarter of the recording’s length, and less on the Fast model than on Accurate.