Hindi with English numbers
Hindi sentences may contain English numbers and technical terms: “पोर्ट four forty three मत खोलो” is a Hindi use case.
Choose language=hi; no separate mixed-language mode or profile is needed. The original URL is retained for compatibility.
| Need | API |
|---|---|
| Hindi recording to text | gpu-asrhq transcription with language=hi (dedicated Hindi recognizer: no punctuation, English words in Devanagari, numbers as digits) |
| Hindi recording with English terms in Latin script and punctuation | Same endpoint with model=asr-legacy, language=hi (general recognizer) |
| Hindi recording to English | Transcribe with model=asr-accuracy, then send the transcript to /v1/translate with source=hi, target=en |
| Live Hindi captions | gpu-live with source_lang=hi |
| Realtime Hindi text with commit control | gpu-asr, pin the Hindi language |
| Hindi text to English | POST /v1/translate with source=hi, target=en |
| Spoken numbers to digits | POST /v1/normalize with language=hi |
| Hindi text to speech | gpu-tts with "language": "hi" — 20 voices, numbers read as Hindi words (text-to-speech) |
The Hindi accuracy recognizer stays loaded and is used only for explicit language=hi recordings. Its text has no punctuation and
writes English words in Devanagari.
The accuracy file option returns raw_text alongside formatted text. Spoken English numbers in Latin or Devanagari,
and recognized Hindi cardinals (“उन्नीस सौ तिरसठ” → 1963, “एक लाख बीस हज़ार” → 120000), can be formatted as digits. Formatting cannot repair an incorrectly recognized number.
Review ports, IPs, leading zeros, negations and corrections before acting on a transcript.
The legacy model=asr-legacy, task=translate option remains available for compatibility. For Hindi, prefer the separate
accuracy transcription and text translation steps, keeping both outputs visible; the legacy recognizer can mishear numbers.
The live API itself returns original-language captions. The dashboard’s optional translation control sends each finalized caption to the separate text translation API and displays both outputs. It does not enable the paused speech-to-speech host.
Public recordings and short smoke tests establish bounded evidence; they do not prove accuracy for your classroom audio. See Speech-to-text, Text-to-speech, Translation.