Bodhan AI Indic Models: OCR, Translation & Speech for 22 Indian Languages
Try this once. Open your favourite AI chatbot and type a normal Hinglish sentence, the way you would text a friend. Something like "kal ka physics test 10 baje hai, chapter 3 revise kar lena". Now ask it to read that out loud, or to translate a scanned page of your Hindi textbook, or to turn a Tamil lecture recording into notes.
For a lot of us, this is where the magic starts to wobble. Most big AI tools are built English-first, and Indian languages get treated as an add-on. That is why a news item from early September caught my attention more than most launch announcements this month.
On September 4, 2026, Bodhan AI, an IIT Madras-incubated centre of excellence for AI in education, released four open models for Indian languages together with AI4Bharat. They cover reading scanned pages, translating, transcribing speech and speaking text aloud. I have not tested these myself yet, so in this post I will stick to what has been published, point out where the claims are the makers' own, and tell you how you could try them cheaply.
The short version
- What it is: four separate models for OCR, translation, speech-to-text and text-to-speech across Indian languages.
- Who made it: Bodhan AI (IIT Madras) with AI4Bharat, part of the Bharat EduAI Stack, which is meant as public digital infrastructure for education.
- How you get it: open weights you can download, or a hosted API priced in rupees, and it is very cheap.
- Who it is for: developers, edtech builders, researchers, and students who want to build something with Indian-language content.
- The catch: the benchmark numbers are the team's own, and there are real limits, like a 30-second audio window.
The four models at a glance
| Model | What it does | Size |
|---|---|---|
| IndicOCR | Turns a page image into structured text | 33M layout model + 0.8B OCR model |
| Indic-Translate | Translates text and whole documents | 4B effective parameters, 32K context |
| Indic-Transcribe | Speech to text (two versions, Core and Flex) | 1.2B each |
| Indic-Speak | Text to speech with 45 voices | about 3.36B |
1. IndicOCR: reading scanned pages
IndicOCR reads printed text in English and all 22 scheduled Indian languages across 13 scripts. It also handles handwriting in English and 12 Indian languages, including Hindi, Bengali, Tamil, Telugu and Urdu. It works in two steps: one small model finds the blocks on a page and their reading order, and a second model reads each block. Equations come out as LaTeX and tables keep their structure.
The team reports a score of 92.76 on OmniDocBench v1.6 (English subset) and 86.2% word-level accuracy on its own internal test across the 22 languages. Treat those with care. An English document score does not prove equal accuracy in every Indian language, and the internal test is the team's own. The makers also flag dense reading order and difficult handwriting as weak spots.
Where I would use it: digitising old textbooks, making scanned notes searchable, or preparing pages for a chat-with-your-PDF setup like the one in my guide on chatting with your own PDFs.
2. Indic-Translate: whole documents, not just sentences
Indic-Translate is a fine-tune of Gemma 4 E4B IT that supports English and all 22 scheduled languages in both directions. Its selling point is document-level translation. It is trained to keep Markdown, LaTeX, tables and code intact while translating the text around them, and it copes with Romanised and code-mixed input.
On the team's in-house document test, it scores 58.97 dBLEU, against 47.44 for Sarvam Translate and 31.93 for IndicTrans2-1B. Higher is better. Human evaluation was still in progress at launch, so read this as a promising early result, not a final ranking.
Good to know: translation between two Indian languages currently goes through English, and direct pairs are on the roadmap. The 32K context also means you cannot feed it an endless PDF in one go.
3. Indic-Transcribe: speech to text, in two flavours
Indic-Transcribe covers the 22 scheduled languages plus English, Bhili and Bhojpuri. Core aims for accurate native-script transcripts. Flex can output native script, Romanised text or a mix, where English words and numbers stay in Latin letters, which is how many of us actually write. The team reports error scores of 8.7 for Core and 11.1 for Flex on its Voice of India test across 15 languages, and says it trained on about 1.3 million hours of audio. Lower is better, and these are again the team's own figures.
The big practical limit is that audio is handled in windows of up to 30 seconds. A long lecture needs to be chopped into chunks first. Live streaming and separating overlapping speakers are listed as future work.
4. Indic-Speak: text to speech that handles code-mixing
Indic-Speak speaks 22 Indian languages across 12 scripts with 45 voices. The part I like most is that it accepts native script and Latin script in the same sentence without you tagging each word's language, which suits how Indians really write and speak. A text normaliser also turns numbers, dates and notation into spoken form.
In the team's test of 30,000 readings, about 93% landed in the top scoring band. That measures whether the audio matches the content, not how natural it sounds to a human ear. Listening tests were still in progress. The makers also admit that quality varies by voice and that some outputs repeat or skip words.
How to access them and what they cost
You have two routes. You can download the weights from Hugging Face and run them yourself, or use the hosted API, which follows an OpenAI-style request format with keys issued per model. Hosted prices are listed in rupees:
| Model | Hosted price |
|---|---|
| IndicOCR | ₹0.20 per image |
| Indic-Translate | ₹0.20 per 10,000 output tokens |
| Indic-Transcribe | ₹0.10 per minute of audio |
| Indic-Speak | ₹6 per 10,000 characters |
To make that feel real: scanning a 100-page book with IndicOCR would cost about ₹20, and transcribing an hour of audio about ₹6. New accounts were listed with ₹10 of starting credit, which is enough to try each model on a small sample. Prices and credits can change, so check the Bodhan console before you plan anything.
What could a student actually build?
Here is a workflow that made sense to me as a B.Tech student. Photograph your handwritten or printed notes and run them through IndicOCR. Translate the verified text with Indic-Translate if your study material is in another language. Turn the result into audio with Indic-Speak so you can revise while travelling. Record a lecture and use Indic-Transcribe for searchable notes. None of this is a ready-made app, so you would connect the pieces yourself, but it is a great weekend project, and a strong one for a portfolio.
If you want to pair this with better study habits, my post on studying with AI for B.Tech exams covers the workflow side.
My honest take
I like the direction. India has spent years relying on models that treat our languages as a bonus feature, and releasing open weights plus cheap rupee pricing lets students and small teams build without a big cloud bill. Bodhan calls these digital public goods, and that framing matters more to me than any single benchmark score.
But I would not treat the launch charts as gospel. Start small. Take one page or one recording in your own language, run it through, and read the output carefully. If it holds up on your real material, then scale up. A quick test I would try: write a Hinglish sentence with numbers and English words in it, run it through Indic-Speak, and check how it handles the mix.
Frequently asked questions
What are Bodhan AI's Indic models?
They are four open models released in September 2026 by Bodhan AI and AI4Bharat: IndicOCR for reading documents, Indic-Translate for translation, Indic-Transcribe for speech recognition and Indic-Speak for speech generation.
Who created Bodhan AI?
Bodhan AI is a centre of excellence for AI in education incubated at IIT Madras. It built the models with AI4Bharat, the IIT Madras lab known for open datasets and tools for Indian languages.
Which Indian languages do they support?
Most of the models cover English and all 22 scheduled Indian languages. Handwriting recognition currently covers 12 Indian languages plus English, and Indic-Transcribe also lists Bhili and Bhojpuri.
Are the Bodhan AI models free?
The weights are open, so you can download and run them yourself if you have suitable hardware. The hosted API is paid but priced very low in rupees, and new accounts were listed with a small starting credit.
Can I run them on my own computer?
Yes, the weights are published on Hugging Face. The models are relatively small, but you will still need suitable hardware and some technical setup. If you are new to running models locally, start with the hosted API for a first test.
Is Indic-Translate better than Sarvam Translate or IndicTrans2?
On the team's own document-level test it scored higher than both. That is the makers' evaluation, and human evaluation was still pending at launch, so test it on your own content before you decide.
Can Indic-Transcribe handle long recordings?
Not in one go. Audio is processed in windows of up to 30 seconds, so long lectures or interviews need to be split into chunks first.
Does Indic-Speak support Hinglish and mixed languages?
Yes. It accepts native and Latin script in the same sentence without you tagging each language. Quality still varies by voice, so listen to a few voices before choosing one.
Sources and where to verify
- Business Standard: Bodhan AI launch coverage
- The Week: multilingual AI stack from IIT Madras
- Analytics Vidhya: model-by-model breakdown
Last checked: September 24, 2026. Benchmarks are reported by the model makers. Prices and access terms may change, so confirm on the official Bodhan AI console.