Sarvam: Speak the language people speak.
Sarvam builds AI models in India for Indian languages, designed for voice first. Explore when meeting people in their own language and by voice opens a market that a translated text app cannot, and what it costs to build.
See the idea in one picture ↓- Indian languages, 2026 models
- 22
- first, not text first
- Voice
- chosen by the IndiaAI Mission
- 2025
- languages in Bulbul’s voice
- 11
Figures as reported by the company.
Independent educational analysis. Scenarios, people, numbers and illustrations on this page are original PMcademy practice material, not company research results or an endorsement.
From typing in English to speaking your language
- 1A text app
Reading and typing in English leaves out people who do neither comfortably.
- 2Voice first
A model built in India hears the question in the person’s language and answers out loud.
- 3Many languages
The 2026 models are described as covering all 22 scheduled Indian languages.
Animated illustration. Names, numbers and people are fictional teaching material.
Most software assumes people read and type.
Read the screen, type in English
Speak in your own language, hear the answer
Read the analysis
Imagine a service that could help millions of people, but only if they can read the screen and type a question. Many people speak several languages fluently and type in none of them comfortably. Translating the screens helps some. Letting people simply speak, in the language they use at home, changes who can use it at all. The product question is who your interface quietly leaves out.
Change the way in, and the market changes.
Illustrative data and simplified behaviour. This model does not represent a live company system or measured company results.
Read the analysis
Use the model below to compare three ways into the same service. An English text app, the same app translated, or a voice assistant that understands and answers in the person’s own language. Watch how many of ten fictional villagers finish a simple task. The numbers in the model are illustrative.
Think about the person who explains things out loud.
“I can explain it in five sentences. I cannot type it in English.”
- To be understood as she speaks
- An answer she can hear
- English-only screens
- Forms she has to read
Read the analysis
Our fictional learner, Lakshmi, runs a dairy co-op in a Telugu-speaking village. She can describe a problem with a payment in five clear sentences. She cannot type them in English, and a translated form still asks her to read and choose. Ask how your person naturally explains a problem, and whether your product can listen in that way.
Benchmarks need their numbers beside them.
Sarvam says its models outperform leading models on Indian language benchmarks.
The benchmark figures, who checked them, or whether real tasks get done.
- Task success by language
- Speech errors by dialect
- Cost per conversation
Read the analysis
Sarvam says its models outperform leading models on Indian language benchmarks, but the post does not publish the figures or who checked them. A benchmark also cannot tell you whether a farmer finished a real task. In your project, measure task success for each language with real people, and keep your own test set for every dialect you claim to support.
Building locally is slower and more expensive.
Read the analysis
Training and running models for many languages, accents and dialects costs far more than adding a translation layer, and each one needs its own testing. Local models may also trail the largest global models on general tasks. In return, the product reaches people others miss, and data can stay in the country. Decide which languages and tasks matter most, and how you will know each one works.
Translate, adapt or build for voice?
Your fictional team runs a state health helpline that today answers in two languages. You can add machine translation to the current text chat, adapt an existing open model for six more languages, or launch a voice assistant built for local languages that hands over to a nurse when unsure. Choose one and explain how you will test each language.
Choose a next move
Leave with a language and voice decision.
Your artifact
Write a language and voice decision: the languages you are missing, translate, adapt or build, whether voice matters, the test task per language and the person who takes over.
Critique my project artifact using only the evidence I provide. Write a language and voice decision: the languages you are missing, translate, adapt or build, whether voice matters, the test task per language and the person who takes over. Separate observed facts, assumptions, proposed tests and missing evidence. Identify a failure mode and a cheap check. Do not invent company results or user research. Ask for my artifact before starting.
Read the analysis
List the languages your users speak at home and the ones your product supports today. Pick the group you are missing most. Decide whether to translate, adapt or build for them, and whether voice matters. Write the task you will test in each language, the success rate that counts as working, and the person who takes over when the system is unsure.