Blog · Indian languages

Speech AI for India has to handle code-mixing, not just 22 languages

Hundreds of millions of Indians speak more than one language, often in one sentence. The open datasets and models are catching up; one setting can still move error from 40% to 6%.

10 October 2026 · 2 min read · Soham Mitra

The shape of the problem

India's 2011 Census counted about 26% of the population as bilingual and 7% as trilingual, as summarised here. In a meeting that becomes code-mixing: a Hindi sentence with English nouns, verbs and dates in the middle. A model that handles 22 languages one at a time can still fail the most common meeting in the country.

Open work worth knowing

• IndicVoices (AI4Bharat) released 7,348 hours from 16,237 speakers across 145 districts and 22 languages, 74% of it extempore speech, under CC-BY-4.0.
• IndicConformer, a 600M-parameter model covering all 22 official languages, MIT licensed.
• Vistaar introduced IndicWhisper, trained on 10.7K hours in 12 languages, with the lowest error in 39 of 59 benchmarks.
• MUCS 2021 set the code-switching benchmark: about 600 hours, with Hindi-English and Bengali-English pairs and baselines above 30% word error.

Commercial work is moving too. Sarvam reports about 19% word error on IndicVoices for Saaras V3 across 22 languages plus English (a vendor claim), and released 30B and 105B models under Apache 2.0.

What we measured

In our six-meeting benchmark, the Whisper models we ship today were far behind on Indian languages. The single biggest swing was not a model at all: Saaras in its default mode wrote English words in Devanagari and scored 39.7% on Hinglish; in its code-mix mode, 6.3%.

What it means for builders

• Test on code-mixed speech, not only on clean single-language sets.
• Score code-mixed output against natural mixed-script text, or the numbers will punish the right answer.
• Do not claim a language until it is measured on real meetings. We have not yet, and say so.