Anuvaya Labs New Delhi, IND
soma-asr-1 · developer preview

Hinglish that comes back romanised

The ears for the agent you are building. Transcripts come back in the Latin script, which is the form an agent reads Hinglish best in, and they stay accurate across the accents people actually speak in.

The sample
haan haan main sun raha hoon... bas ek second, network thoda weak hai. haan, toh main keh raha tha ki wo wala plan zyada better hai.

One clip, transcribed by Soma. Hesitation, filler, and a mid-sentence switch, returned as romanised text and ready to hand to whatever reads it next.

Romanised

Latin script, because that is what reads it next

Almost nothing that reads a transcript is a person any more. An agent handles Hinglish better in Latin script than in Devanagari, so romanised is what comes back, rather than a setting you have to go and find.
  • haan ji bataiye, aapke order mein kya dikkat aa rahi hai?

  • aapki appointment kal shaam paanch baje hai, confirm kar dein na?

  • achha ek kaam kijiye, app mein jaake status check kar lijiye.

Accents

The country does not speak in one accent

Speech models trained mostly on one region's English get quietly worse the further you travel from it. That is the failure this model is built to avoid.

Understood across the country.

Regional English and regional Hinglish both land, without the speaker having to flatten how they talk to be transcribed correctly.

  • Delhi and the Hindi belt
  • Mumbai and the west
  • Kolkata and the east
  • Chennai, Bengaluru, Hyderabad, and the south
  • Punjab and the north
  • The Northeast
Capabilities

What that means in practice

Romanised, not transliterated after the fact
Romanised Hinglish is what the model returns, not Devanagari run through a converter, so it comes back spelled the way people actually type it.
Code-switching mid-sentence
Move from English to Hindi and back inside one sentence. The transcript follows without breaking the thought.
Accents from across the country
The same audio does not need a Delhi accent to become useful text. Regional English and regional Hinglish both land.
Everyday recordings
Phone calls, laptop mics, and a room with other people in it. These are the conditions most speech actually gets recorded in.

Also transcribes

Hinglish (romanised), Hindi, Bengali, Tamil, Telugu, Marathi, Gujarati, Kannada

Try it on your own audio

soma-asr-1 is in developer preview. It is early, and changing quickly.

Part of Soma, a new family of TTS and ASR models designed for conversational AI in India. For the voice, see soma-tts-1. Built by Anuvaya, an applied AI lab making AI conversational and human-like for India.