Iyal AI
Language Technology Research Lab

A Sovereign Lab for
Tamil Intelligence.

Speech recognition, neural voices, translation, writing tools, document intelligence, a conversational assistant and adaptive learning — built on our own Tamil corpus, tokenizer and evaluation set, and live on this page right now.

Use the ToolsLive and free to useThanjavur, Tamil Nadu
LIVE PRODUCTS
08 In Public Use
OUR TAMIL TOKENIZER
5.4× Fewer Tokens
TAMIL VOICES
IN · LK · MY · SG
Live Products

எட்டு தமிழ்க் கருவிகள். இப்போதே பயன்படுத்துங்கள்.

Eight Tamil products,
running right now.

Each one has a name, a job and a working interface. Pick one and use it — no sign-up, no waiting list, nothing here is a mock-up.

பேசு PesuTamil speech to text

Speak Tamil, see Tamil.

Press the mic and speak in Tamil. Your words appear as Tamil text while you talk, and each finalised segment is kept with a timestamp.

Tamil layer by Iyal AI: Tamil variety handling and the held-out Tamil evaluation set

திற · Open the full product
  • Tamil (India)
  • Writes as you speak
  • Free to use
Tamil variety

Tamil — India

Ready
Try saying:“வணக்கம், நான் தமிழ் பேசுகிறேன்”“இன்று வானிலை நன்றாக இருக்கிறது”“தஞ்சாவூர் பெரிய கோவில்”

Press “Start speaking” and speak Tamil into your microphone.

Your words will stream here in real time as native Tamil Unicode text.

0 words·0 letters

Iyal AI builds the Tamil layer: the corpus, the tokenizer, the morphological analyser, the pronunciation scoring and the evaluation set. Commodity acoustic and translation engines are integrated underneath where buying beats building, and Tamil text analysis runs entirely in your browser.

Roadmap

தமிழுக்காக உருவாக்கப்பட்ட AI. எங்கள் சொந்த முயற்சி.

Iyal-Tamil-2B
a Tamil model of our own.

The Tamil corpus, the tokenizer and the evaluation set behind these products are ours and already measured. The next step is the model itself: a Tamil-first language model we train, evaluate and licence ourselves — two billion parameters, small enough to run on a phone.

Phases 00 and 01 have happened, and every number below them is measured rather than projected. Phases 02 onwards are planned work; timings assume funded compute.

  1. 00

    Products in public use

    LiveLive today

    The Tamil products above, across speech, language and learning. They also produce the held-out Tamil test material and the evaluation harness every later phase is measured against.

    • 8 live products
    • Tamil evaluation harness
  2. 01

    Corpus & tokenizer

    In progressWeeks 1–6

    Collect, clean and deduplicate Tamil text with provenance recorded per source, then build a tokenizer so Tamil words stop fragmenting into meaningless pieces. A pilot run over 67,587 Tamil Wikipedia documents (16.3M words) is complete; the result below is from it.

    Tokens needed to read the same Tamil
    • Iyal Tamil tokenizer1.81 tokens per word
    • BLOOM (multilingual)2.21 1.2× more
    • Qwen2.5 (open 1.5B)9.77 5.4× more
    • GPT-2 (English)24.18 13.3× more

    Fertility measured over 300 held-out Tamil documents. Lower is better — fewer tokens means lower cost, longer context and faster replies.

    • 67,587 documents cleaned
    • 32k Tamil tokenizer trained
    • Fertility measured
  3. 02

    Continued pretraining

    PlannedWeeks 7–10

    Continue training a permissively licensed 2B base on the Tamil corpus, so Tamil stops being the degraded case. Measured from the pilot: roughly 250 GPU-hours for 30 billion tokens seen — about 31 hours on eight GPUs.

    • Iyal-Tamil-2B base checkpoint
  4. 03

    Instruction tuning

    PlannedWeeks 11–13

    Teach the model to follow instructions and hold a conversation in Tamil across registers, from spoken everyday Tamil to formal written Tamil.

    • 50–100k Tamil instruction pairs
    • Chat checkpoint
  5. 04

    Evaluation & release

    PlannedWeeks 14–16

    Measure on held-out Tamil benchmarks, publish the results including where the model is weak, and release with ownership and licensing stated up front.

    • Published benchmark results
    • Model card & licence
    • Quantised build for phones
Our Own Work

எங்களுடையது எது? சரிபார்க்கும் வழி என்ன?

What is ours,
and how to check it.

Five claims. Each is stated so it can be tested rather than taken on trust, and the first two are already measured and reproducible on request.

  • 01Measured

    A Tamil tokenizer built around Tamil morphology

    General tokenizers cut Tamil statistically, as a character sequence, which shatters inflected words into fragments that mean nothing. Ours is built with Tamil word formation in mind — roots and productive suffixes as units — so Tamil costs less to read, fits further into context and comes back faster.

    Evidence: Measured: 1.81 tokens per word against 9.77 for Qwen2.5, over 300 held-out Tamil documents — 5.4× fewer tokens for the same Tamil. Re-runnable with a single command.

  • 02Measured

    An open Tamil benchmark we can be beaten on

    Without a public Tamil benchmark, every claim of Tamil AI capability is unfalsifiable. We built one — cloze items generated from held-out documents, plus hand-written grammar items on sandhi doubling, case endings, agreement, plurals and word order — and we publish our own scores on it, including where we do badly.

    Evidence: 212 items built and scored. A strong general model reaches 72.6% on Tamil a school student would find easy; chance is 19.8%. That gap is the work.

  • 03Scoped

    Dialect-balanced Tamil speech, reported per dialect

    Public Tamil speech systems are trained mostly on read, formal, urban speech, then deployed on a helpline where someone speaks quickly in a regional variety with English words mixed in. We work across the Indian, Sri Lankan, Malaysian and Singaporean varieties and intend to report accuracy per variety rather than as one flattering average.

    Evidence: Four Tamil varieties are live in recognition and in eight neural voices today. Per-dialect error rates on spontaneous code-mixed speech are not yet measured; the evaluation harness for it is built.

  • 04Scoped

    Recognition of the pre-1978 Tamil script

    The Tamil script was reformed in 1978. Much of the twentieth-century Tamil literary and periodical record uses the older orthography, which current recognition handles poorly — so a large archive sits unreadable by machine. There is no commercial champion for fixing it and clear public value in doing so.

    Evidence: Scoped and designed, not yet built. Named here because it is a specific, checkable commitment rather than a general ambition.

  • 05Scoped

    Tamil personal-data detection, and the means to measure it

    Anonymisation tools built for English do not reliably catch Tamil personal names, Tamil place names or Indian identifier formats written in Tamil script. Reliable anonymisation is a precondition for sharing Tamil data lawfully at all, so this is not merely a compliance item — it is what unlocks the corpus scarcity the whole field is stuck behind.

    Evidence: The corpus pipeline already records provenance and licence metadata for every document. Tamil personal-data detection and its benchmark are scoped and not yet built.

Our Projects

Built by Iyal AI,
running in the real world.

Eight Tamil products, each live and free to use. Built on the corpus, tokenizer and evaluation set below — open any of them and try it.

Contact

Let’s Build
Tamil’s AI Future

Fill in the form below

Add an Attachment