Zhagaram Labs Lend your voice

Why Zhagaram Labs exists

More than 80 million people speak Tamil. It’s one of the world’s oldest living languages, with an unbroken literature of more than two thousand years. Yet the AI tools most people now rely on were built for English first, and Tamil comes out badly. They mishear it, charge more to process it and mispronounce its numbers. They also can’t search it, spell-check it or read its old books.

Zhagaram Labs is a community, non-profit lab. We exist so that Tamil has its own open models and tools, good enough to build the best products on, and free for anyone to use. Every claim we make comes with a number anyone can check.

Here are the problems we’re solving, one at a time.

For people who talk rather than type

1 A grandmother calls to ask about her pension

What happens today

She doesn’t type and doesn’t read English, so a phone call in her own Tamil is her best way in. But speech recognition still gets much of Tamil wrong.

The evidence

Whisper large-v3-turbo, a widely used open speech model, gets about two in three Tamil words wrong (67% word error rate) on clear read speech. That’s on both the FLEURS and Kathbath test sets. Measured by us, Uraikal track A1. Real calls are harder: phone audio, background noise and regional speech.

What we’re building

  • Sevi, speech recognition trained for dialects, phone audio and Tanglish.
  • Kural, a voice assistant you reach with an ordinary phone call. Building

2 A farmer asks about a loan at “ரெண்டு ரூபா வட்டி”

What happens today

In Tamil Nadu, “2 rupees interest” means 2% a month, which is 24% a year. A model that reads it as 2% a year is out by 12 times on a farmer’s loan. Tamil number words are also fused and irregular. 90 is தொண்ணூறு and 900 is தொள்ளாயிரம். 1.5 lakh is ஒன்றரை லட்சம்.

What we’re building

  • Exact, tested number handling in both directions, from Tamil words to digits and back. Released in uraikal.
  • Tamil number and unit reasoning for our language model. Building

3 A voice reads out a price, a date or a phone number

What happens today

Before a voice can speak “₹1,250” or “2024ல்”, text has to be turned into the words a person would say. NVIDIA’s open toolkit for this has no Tamil at all. The one open Tamil extension has open bugs on basics: it writes 300 as “முன்நூறு”, reads dates as phone numbers, and leaves “2024ல்” as digits. Source: NeMo text processing; Kenpath issue #4.

What we’re building

  • Tamil text normalisation for speech, tested item by item. Building
  • Oosai, our Tamil voice. Building

4 A speaker from Jaffna, Kongu or Tirunelveli

What happens today

Tamil changes every few hundred kilometres: Chennai, Kongu, Madurai, the Cauvery delta, Tirunelveli and Kanyakumari, Jaffna, Batticaloa, Malaysia, Singapore. No open test measures speech recognition region by region, so nobody knows who gets left behind.

What we’re building

  • Test sets for nine regions, recorded with consent.
  • A headline score that averages across regions, so no system can top our leaderboard on clean Chennai-style speech alone. Planned

5 Nobody talks like a newsreader

What happens today

Written Tamil and spoken Tamil differ in sounds, verb endings and case endings (போகிறேன் / போறேன்). Models trained on news and the web answer in formal written Tamil. That sounds stiff, even cold, to the people we most want to reach.

What we’re building

Models that understand both kinds of Tamil and answer in warm, spoken Tamil, with a test that checks they do. Building

6 “Login பண்ணுங்க”

What happens today

Many Tamil speakers mix English words in, and type Tamil in English letters (Tanglish). The same word comes out as poren, pooren or porean. Few tools treat this as normal Tamil.

What we’re building

Tanglish test sets, and conversion between Tanglish and Tamil script. Planned

For people building with AI

7 Tamil costs more on every AI model

What happens today

AI models read text as “tokens”, and you pay per token. Many tokenizers split Tamil letters apart: they cut a consonant from its vowel sign before they even start. Qwen3, a popular open model family, cuts 98.8% of Tamil words this way. It needs 9.99 tokens per Tamil word, against 2.44 for Gemma 3 and 1.68 for the Tamil-trained Agaram. So Tamil costs about 4 times more on Qwen3 than on Gemma 3, fits less into the model’s context window, and runs slower. Measured by us, Uraikal track D4, 13 tokenizers.

What we’re building

  • A public tokenizer leaderboard. Released
  • A one-line fix for the splitting bug.
  • Tamil tokenizer extensions for popular open models. Building

8 Search for நாடு and miss நாட்டில்

What happens today

Tamil packs a whole phrase into one word, and letters change where word parts meet (நாடு + இல் → நாட்டில்). So a search engine has to understand word structure to find anything. Elasticsearch, one of the most widely used search engines, has no Tamil analyzer. The only Tamil search test we found is machine-translated from English. Source: Elasticsearch language analyzers.

What we’re building

  • Thedal, Tamil search: an analyzer for common search engines, a small search model, and the first search test written in original Tamil. Planned

9 A spell checker that flags correct Tamil

What happens today

No open Tamil spell or grammar checker publishes how often it’s right. The open grammar checker LanguageTool has 210 Tamil rules and no spell check. The most-used Tamil checker’s core is reportedly closed. Word-list checkers fail on Tamil, because one root can take thousands of correct forms. Source: LanguageTool languages.

What we’re building

An open word-structure analyser, a checker for sandhi (letter changes where words join), and a spell and grammar checker. Each comes with a published accuracy. Planned

10 Pronunciation follows rules that tools ignore

What happens today

One Tamil letter can sound different depending on where it sits. க is “k” in கடல், “g” in தங்கம் and close to “h” in மகன். Epitran, a common open pronunciation tool, has no rules for this. Source: Epitran Tamil rules.

What we’re building

A Tamil pronunciation library, checked by a linguist against a word list. Building

11 Is this Tamil simple enough for a child, or for a first-time reader?

What happens today

English has readability scores. For Tamil we found no open readability measure at all. Without one, a model can’t check that its answer is simple enough, and nobody can sort learning material by reading level.

What we’re building

A Tamil readability measure, built with teachers. Planned

For Tamil’s heritage

12 Old books, government orders and palm leaves

What happens today

  • Tesseract, the standard free OCR tool, gets 61.5% of words wrong on printed Tamil, and its Tamil model hasn’t changed since 2017. Source: Jayatilleke & de Silva, 2025.
  • Many government PDFs use pre-Unicode Tamil fonts, so copied text comes out as gibberish.
  • For palm-leaf manuscripts, there’s no released model at all.

What we’re building

  • Olai, for reading Tamil from images.
  • A converter for old Tamil fonts. Planned

13 The Thirukkural, quoted wrong

What happens today

Language models can make up verses and present them as genuine. Tamil prosody (யாப்பு) can be checked by computer. But the existing tools publish no accuracy, and the best known has no licence that allows reuse.

What we’re building

  • Exact quotes from trusted texts, never from model memory.
  • A prosody checker tested on all 1,330 Kurals. Building
  • A test of whether popular models know the Kural. Planned

Underneath all of it: you can’t improve what you don’t measure

In most of these areas there’s no Tamil test set at all, so nobody can say how good the tools are. So we built Uraikal (உரைகல், “touchstone”): an open benchmark and toolkit. Every model and tool we release has to beat the last one on it.

  • Uraikal is released: pip install uraikal, on GitHub under Apache 2.0.
  • What it scores today: speech recognition and tokenizers.
  • What’s coming: prosody, speech text normalisation, pronunciation and more.

How we work

  • Open by default:
    • Code under Apache 2.0.
    • Models on Hugging Face.
    • Data under CC BY 4.0, where the people who gave it agree.
  • Consent first: we don’t take data without clear consent, and we don’t clone voices.
  • Safety: Kural never diagnoses, never asks for OTPs or bank details, and says “I don’t know” rather than guess.
  • Tamil only, done properly: other labs cover many Indian languages. We go deep on one.

Help us

Record your Tamil, check our examples if you’re a Tamil teacher or linguist, or build on what we release.

Help a machine hear your Tamil

We need speakers from Chennai to Kanyakumari and Jaffna to Kuala Lumpur, of every age. Tell us where you grew up and we'll write when recording opens.

Lend your voice

Opens an email to us. Nothing is recorded without your consent.