Home › Hard problems › India

India: hard problems at India's scale

Speech recognition for dialects and code-mixed talk

Solo: strongOne person can run a real test of this in about 8 weeks.

"Bhai, kal market chalein?Traffic bahut hai yaar"Hinglish, with aBhojpuri accenthonk!street noisebad phone lineSpeech modeltrained onstandard Hindiwhat it writes downBhai, kalmarket ka lentrophybahuthaiyathe words that matter come out wrong(the red ones, as an example)
Speech models learn standard Hindi, but people talk in dialects, switch to English mid-sentence, and call from noisy streets.

Why it is still open

Models trained on standard Hindi fail on Bhojpuri, Marwari, Hinglish switching mid-sentence, and noisy streets and phone lines.

An experiment one person can run in 8 weeks

Record or collect 500 clips across 5 dialects. Measure the word error rate of open models (Whisper, AI4Bharat's models) and test a small fine-tune.

Verdict: Solo: strong

Open models exist and the error rate is a clear number.

The solo test behind the verdict

  1. Can it be measured? There is a number that shows you did better, such as an error rate, a catch rate or a cost per answer.
  2. Is it cheap to try? It runs on a laptop, free open models and open data, or under about ₹50,000 of rented GPU time.
  3. Is there a public yardstick? A benchmark or dataset others use, so your result can be compared and believed.
  4. Would a small win get noticed? A lab, a funder or a customer would care about a 10% improvement, not only a full solution.

Four yeses means Solo: strong. Two or three means Solo: possible. Fewer means Follow, don't build.

PreviousAI in Indian languages costs more and works worseNextReading Indian documents: land records, forms, handwriting

All 79 hard problems · Working on this, or know of a result we missed? basant@cleandataengineering.ai