📖 Chapter 3 of 9 · AI For New Learner Advance 2 (Multi Indian Language) · Class 8
Chapter 3: AI That Speaks Your Language
In this chapter
- Why English-only AI was a problem for India.
- Bhashini — the government’s gift to 1.4 billion Indians.
- How AI understands language: words to numbers to meaning.
- Sarvam AI, Krutrim, BharatGPT — India’s own LLMs.
- What ‘low-resource language’ means and why it matters.
The Big Idea
India has 22 official languages and over 1,600 mother tongues. For decades, the smartest AI only worked in English — leaving most Indians out. That is now changing fast. The Government of India, IIT labs, and Indian startups are building AI that listens, reads, and writes in YOUR language. This is one of the most exciting AI revolutions on Earth.
Did you know?
Only about 11% of Indians can read English fluently. Until 2022, almost ALL major AI was English-only. Today, Bhashini and Indian LLMs are closing that gap — and India is leading the world in multilingual AI research.
Story 1: Saanvi Brings Maithili to AI
From: Patna, Bihar · Saanvi, age 12
Saanvi speaks Maithili at home with her Naani in a Madhubani village. At school in Patna, everything is in Hindi and English. Last week she got a strange homework: translate a Naani’s story from Maithili to English.
She tried Google. Most results were in Hindi. Then her cousin showed her Bhashini — the government’s free AI translator. Saanvi spoke the Maithili story into her phone. In seconds, Bhashini wrote it in English. Saanvi was AMAZED.
Even better, Bhashini saved her recording. Saanvi later learned that her voice is now part of the data that will teach FUTURE AI to understand Maithili even better. “My Naani’s words are training the AI,” Saanvi whispered, proud.
Made-in-India language AI
Real products serving real Indians today.
📁 CASE FILE Bhashini — Govt of India
Launched: 2022 by Ministry of Electronics & IT.
Languages: Translates between 22 official Indian languages.
How: Voice in, voice out, text in any direction.
Used by: Farmers, railways, courts, healthcare.
Open: Free for all citizens and built on open data.
📁 CASE FILE Sarvam AI — Bengaluru startup
Founded: 2023 by ex-IIT and Google researchers.
What: Built Indian LLMs that ‘think’ in Indian languages.
Scale: Sarvam-1 trained on 2 trillion Indian words.
Used by: Banks and government for voice helplines.
Proof: India can build world-class AI from scratch.
📁 CASE FILE Krutrim — Ola’s Indian LLM
Launched: 2023 by Ola founder Bhavish Aggarwal.
First: First Indian LLM to support 10+ Indian languages.
Voice: Built a voice chatbot for Hindi, Tamil, Telugu users.
Trained on: Indian internet, books, government documents.
Aim: To be India’s own ChatGPT alternative.
How does AI read words?
Words become numbers; numbers become meaning. Computers cannot ‘feel’ the word ‘mother’, so AI turns every word into a long list of numbers called a vector. Words that mean similar things get similar numbers — ‘Mother’ and ‘Maa’ end up near each other; ‘Roti’ and ‘bread’ too. This is called an EMBEDDING. Once words are numbers, AI can do MATH on them — and that math is what understanding looks like for a computer.
💻 BEHIND THE CODE How Bhashini Translates
READ Maithili sentence BREAK into words CONVERT each word → vector (numbers) PASS vectors through trained AI brain GET English vectors back CONVERT vectors → English words RETURN translated sentence
Solved illustrations
Understanding language AI step by step
1. What is NLP?
Show answer
Answer: NLP stands for Natural Language Processing — the branch of AI that deals with human language. Bhashini, ChatGPT, voice assistants — all are NLP systems. NLP turns messy human language (slang, typos, accents) into something computers can work with.
2. Why is Hindi or Tamil ‘harder’ for AI than English?
Show answer
Answer: AI learns from data on the internet. The internet has 100x more English than any Indian language. Less data means less training. This is called a ‘low-resource’ problem. Bhashini and Sarvam AI are solving it by deliberately gathering Indian-language data.
3. What is an ‘embedding’ in simple words?
Show answer
Answer: It is a list of numbers that represent the meaning of a word. Words with similar meaning get similar embeddings. This lets AI do math like: KING − MAN + WOMAN = QUEEN. Yes, AI can do that — and it works in any language!
4. Why is voice AI tricky?
Show answer
Answer: Speech has accents, background noise, half-words, and code-mixing (‘Mujhe paani de do please’). Indian voice AI has to handle all of this. Sarvam AI’s voice models are trained on actual Indian phone calls, so they understand Hinglish and Tanglish.
5. Can I help train Indian language AI?
Show answer
Answer: Yes! Bhashini has a public ‘data donation’ programme. You can read a sentence aloud or label a translation. Every recording helps. Schools across India organise voice-donation drives. Even your voice can build India’s AI.
Questions — Chapter 3
Test what you learnt about language AI
A. Fill in the blanks
- Bhashini is built by the Government of .
- NLP stands for Natural Language .
- AI turns words into lists of numbers called .
- Sarvam AI is a startup based in .
- The Indian LLM launched by Ola’s founder is called .
B. Short answers
- Why did English-only AI leave most Indians out?
- Explain in 2 sentences: how does AI translate Maithili to English?
- What is a ‘low-resource language’ and why does it matter?
- Name two real Indian AI products that help with language.
- How can you personally help train Indian language AI?
Brain Buster
Open Bhashini on a phone. Translate one sentence into 3 Indian languages. Share with class!
Check your understanding — Chapter 3
Tap an answer to check it instantly. These are for practice — nothing is saved or sent.
Q1. What does NLP stand for?
- New Language Program
- Natural Language Processing
- National Learning Plan
Show answer
B. Natural Language Processing
Q2. Why is Hindi or Tamil often 'harder' for AI than English?
- Indian languages are more complex
- There is far less Indian-language data online, so AI gets less training — a 'low-resource' problem
- AI dislikes Indian languages
Show answer
B. There is far less Indian-language data online, so AI gets less training — a 'low-resource' problem
Q3. What is an 'embedding' in simple words?
- A picture of a word
- A list of numbers that represents the meaning of a word, so similar words get similar numbers
- A translation dictionary
Show answer
B. A list of numbers that represents the meaning of a word, so similar words get similar numbers
Q4. How can YOU help train Indian-language AI?
- You can't — only engineers can
- By donating data — reading sentences aloud or labelling translations on Bhashini's public programme
- By buying a new phone
Show answer
B. By donating data — reading sentences aloud or labelling translations on Bhashini's public programme