Zhagaram Labs Lend your voice
Engraving of the Shore Temple at Mamallapuram at sunrise, with palmyra palms on the shore and a kattumaram sailing on the Bay of Bengal.

Speech and language models,
built for Tamil first

Zhagaram Labs builds AI for the more than 80 million people who speak Tamil. We start with Kural, a voice model you can talk to in everyday Tamil, and we release most of what we build as open source.

Lend your voice

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

Your grandmother's Tamil, your cousin's Tanglish, the accent of the town you grew up in. We're building models that understand all of it.

Why Tamil needs its own models

Most Indian language models treat Tamil as one of twenty two languages. That gets you a demo. It doesn't get you a model that talks like someone from Madurai, understands a grandmother in Jaffna, or reads a palm leaf manuscript.

The Tamil letter ழ, zha, engraved in ink lines ழ

ழ, zha, is the last sound in தமிழ். Among Indian languages only Tamil and Malayalam still have it, and it sits beside two close cousins. A Tamil speech model has to hear the difference:

வாழ்
live
வாள்
sword
வால்
tail

In our first benchmark run, Whisper wrote விலையாட்டு for விளையாட்டு, sport. One letter swapped, and the word is gone.

There are two Tamils

Written and spoken Tamil differ in vocabulary and grammar. Both words below mean “I'm going”. Models trained on news text answer like a newsreader.

செல்கிறேன் written போறேன் spoken

It changes every few hundred kilometres

Chennai, Kongu, Madurai, Tirunelveli, Jaffna, Batticaloa, Singapore, Malaysia. Accent and vocabulary shift across all of them, and most benchmarks test none of them.

9 regions our benchmark will score separately

People speak Tanglish

Everyday Tamil mixes in English words and is often typed in Latin script. A useful model hears, reads and writes all of it.

naan office ku late aa poren

One word can be a whole sentence

Tokenizers built for English cut long Tamil words into many pieces, so Tamil runs slower, costs more and fits less into a model's memory.

படித்துக்கொண்டிருந்தேன் one word: I was reading

Kural, our first model

Kural means voice. You speak Tamil and it answers in Tamil as you talk. You can interrupt it, switch to Tanglish halfway through a sentence, or speak the way people speak where you grew up.

  1. Measure it first

    Uraikal, our open benchmark, scores models on spoken Tamil, dialects, Tanglish and the ழ ள ல sounds. Nothing ships unless it beats the last version on it.

  2. Ship a working pipeline

    Sevi listens, Mozhi thinks, Oli speaks. It works early, and every conversation shared with consent improves the training data.

  3. Make it speech native

    Naadi turns audio into tokens, so one model listens and speaks directly, keeps the tone that text throws away, and can listen while it talks.

Models we're building

Kural needs ears, a voice, an audio codec and a Tamil language model. Each is useful on its own, so each ships on its own. Bound here the way Tamil literature was kept for centuries, on palm leaves.

Speech

  • குரல்

    Kural voice

    Talk with it in Tamil and it answers in Tamil as you speak. Handles interruptions, Tanglish and regional speech.

    In progress API first, a smaller open version later

  • செவி

    Sevi ear

    Speech recognition for spoken Tamil in every dialect, including Tanglish and noisy phone audio. Writes Tamil script or Tanglish.

    In progress Open weights

  • ஒலி

    Oli sound

    Speech synthesis that sounds like spoken Tamil, not a newsreader. Reads numbers, dates and English words correctly.

    In progress Open weights, preset voices

  • நாடி

    Naadi pulse

    An audio codec that turns Tamil speech into compact tokens a language model can read and write.

    In progress Open weights

  • தூது

    Thoothu messenger

    Speech translation between Tamil and English, Hindi and other languages, in the speaker's own voice. For dubbing and calls.

    Later API first

  • உச்சரிப்பு

    Uchcharippu pronunciation

    Scores pronunciation of ழ ள ல, ண ன ந and ற ர for learners, schoolchildren and speech therapy.

    Later Open weights

  • வட்டாரம்

    Vattaaram region

    Recognises which dialect someone is speaking, so the other models can adapt to it.

    Later Open weights

Language

  • மொழி

    Mozhi language

    A Tamil language model with a tokenizer built for Tamil, fluent in written and spoken Tamil. The brain inside Kural.

    Next Open weights

  • மாற்று

    Maatru change

    Converts Tanglish to Tamil script and back, and spells out numbers, dates and abbreviations so they can be spoken.

    Next Open weights

  • தேடல்

    Thedal search

    Embeddings for Tamil search that match across Tamil script, Tanglish and dialects.

    Later Open weights

Heritage

  • ஓலை

    Olai palm leaf

    Reads palm leaf manuscripts and old printed Tamil books, including letterforms from before the 1978 script reform.

    Later Open weights

  • யாப்பு

    Yaappu prosody

    Understands classical Tamil: Sangam poetry, the Thirukkural, metre and the traditional commentaries.

    Later Open weights

Data and evaluation

  • உரைகல்

    Uraikal touchstone

    The benchmark we measure everything against: spoken Tamil, every dialect, Tanglish and ழ level pronunciation, with a public leaderboard.

    In progress Open benchmark

  • ஊற்று

    Ootru spring

    Tamil speech and text recorded with consent across regions, released for anyone to train on.

    In progress Open data

Open by default

When we release a model, you can download it, run it on your own hardware and build on it. And we publish our numbers, so you can check them.

What we release

  • Model weights on Hugging Face
  • Training and inference code on GitHub, under Apache 2.0
  • Datasets we have consent to share, under CC BY 4.0
  • Uraikal and its leaderboard, so anyone can check our scores

What we hold back, and why

  • Voice cloning. Oli ships with preset voices, not cloning from a sample, so it can't be used to impersonate someone. Generated audio is watermarked.
  • Kural's full model starts as an API while we test it for misuse. A smaller open version follows.
  • Data we can't redistribute. Licensed data can train a model, but we don't republish it.

What every release includes

  • A model card with Uraikal scores broken down by dialect
  • A data card listing sources, hours and consent
  • The cases where it fails
  • Small builds that run on a laptop or phone

Uraikal, first results

Share of words and characters a system gets wrong on public Tamil test sets. Lower is better. Kathbath's new speakers are people absent from its training data. Our own models join this table as they ship.

SystemTest setClipsWords wrongCharacters wrong
Whisper large v3 turboFLEURS, read speech59167.5%18.5%
Whisper large v3 turboKathbath, known speakers2,75967.6%15.9%
Whisper large v3 turboKathbath, new speakers1,64267.2%16.0%

Questions

Something missing? Write to hello@zhagaramlabs.com.

What does Zhagaram mean?

ழகரம், zhagaram, is the name of the letter ழ. It's the sound at the end of தமிழ், and the one speech models most often get wrong.

When can I talk to Kural?

Kural is in development. The first version chains speech recognition, a Tamil language model and speech synthesis. We'll announce a preview here and on GitHub when it's ready to try.

Which models will be open source?

Most of them. Weights go on Hugging Face and code on GitHub under Apache 2.0. We hold back voice cloning, Kural's full model until we've tested it for misuse, and data we don't have the right to republish.

What happens to my voice if I volunteer?

Nothing is recorded without your consent. Before you record anything, we'll tell you exactly how the audio will be used and whether it will be published in Ootru, our open dataset. You decide.

Which dialects will you cover?

Our benchmark plans nine regions: Chennai, Kongu, Madurai, the Cauvery delta, Tirunelveli and Kanyakumari, Jaffna, Batticaloa, Malaysia and Singapore. Speakers from anywhere else are just as welcome.

Will it understand Tanglish?

Yes, and that's a requirement, not an extra. Sevi writes what you say in Tamil script or in Tanglish, and Maatru converts between the two.

Why not use Whisper or another multilingual model?

We tested it. Whisper gets 67% of words wrong in read Tamil and swaps ள for ல, which changes meaning. Multilingual models spread themselves across dozens of languages. We put everything into one.

Can I test my own model on Uraikal?

Yes. The test sets are public, and the scoring code will be on GitHub under Apache 2.0. Any system that writes one line of text per audio clip can be scored.

Are you hiring?

We're looking for speech researchers, ML engineers and Tamil linguists, full time or as collaborators. Email us about working together.

We hold Tamil recordings or texts. Can we work together?

Yes. Universities, libraries, broadcasters and publishers: we'll help digitise your archive and open it, with credit. Talk to us about data.

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.