Indic TTS normalizer and ITN
Text → spoken (for text-to-speech): paste text with numbers and get the words a voice should say: ₹ amounts, times, dates, years, percentages, decimals, phone numbers, units, ordinals, ranges and cricket scores, in Hindi, Bengali, Tamil, Telugu, Marathi, Gujarati, Kannada, Malayalam, Punjabi and Urdu. Spoken → text (for speech-recognition output) does the reverse: number words become digits, ₹ amounts, times and dates.
Models: altslate/indic-tts-normalizer-small and altslate/indic-itn-small (300M each, Apache-2.0), run in your browser with int8 weights. Your text never leaves your device. Each mode downloads about 300 MB the first time it is used (the optional punctuation step in "Spoken → text", altslate/indic-punctuation-base, another 300 MB); after that a sentence takes a few seconds (the output is written byte by byte). It reads one pattern per format (for example 12/03/2024 is always day/month/year) and was not checked by native speakers; see the model card for results and limitations.