Introduction

Zhagaram Labs — a Tamil-first non-profit AI lab — just open-sourced Kuyil (குயில்), a tiny Tamil text-to-speech family. The Hugging Face cards zhagaramlabs/kuyil-20m, kuyil-5m, and kuyil-1.4m were created 2026-11-04T04:25Z. Code lives at github.com/zhagaramlabs/kuyil.

This is open generative TTS, not a quant of someone else’s Tamil checkpoint. The recipe is explicitly rebuilt from PiVo-1.4M (ArtRealmAI already covered the English PiVo drop) for Tamil script, Latin-script English words, and ₹/number reading.

What Shipped

  • Vendor / product: Kuyil 1.4M / 5M / 20M — Zhagaram Labs
  • Modality: generative text-to-speech (Tamil + Latin English / numbers)
  • Open vs closed: weights CC BY 4.0 (credit Zhagaram Labs + IIT Madras IndicTTS); code Apache-2.0; keep PiVo/FARGAN notices in pivo/
  • Runnable today: clone the repo → uv sync → hf download zhagaramlabs/kuyil-20m → synth.py
  • Concrete fact: on 125 held-out IndicTTS sentences, IndicConformer CER on Kuyil-20M speech is 2.9% — matching the original studio recordings; synthesis ~15× realtime (RTF 0.06–0.07) on one Apple-silicon CPU thread including the vocoder
  • Voices: two IndicTTS studio speakers — female and male

Run It

git clone https://github.com/zhagaramlabs/kuyil && cd kuyil && uv sync
hf download zhagaramlabs/kuyil-20m --local-dir kuyil-20m
uv run python synth.py kuyil-20m "வணக்கம், நான் குயில்." --voice female --out hello.wav

In Python from the repo folder: speak, voices = synth.load("kuyil-20m") then speak(text, "female") returns 16 kHz NumPy audio. Swap the folder name for kuyil-5m or kuyil-1.4m if you want even smaller.

How It Was Built (Short)

  • Stack: text encoder + duration predictor + causal decoder predicting 20 FARGAN features every 10 ms; PiVo’s FARGAN vocoder turns features into audio
  • Data: IIT Madras IndicTTS Tamil (~24.8 h after trim; 9,022 train utterances ≤30 s) via the SPRINGLab mirror
  • Frontend: Tamil letters → 74 phones; Latin English via CMUdict mapped to Tamil phones; numbers spoken the way people say them
  • Training note: ~6.6 h on one H100 (epoch 21 lowest val loss); no PiVo-style adversarial refinement pass

All three sizes are “as intelligible as the recordings” by CER, so pick by taste and footprint — native listeners will hear the difference where a recognizer will not.

Limits and Responsibility

  • Studio literary / read style only — conversational tone and Tanglish speech were not in training (Tanglish text is frontend-only)
  • Naturalness has not been rated by native listeners yet
  • These are real people’s IndicTTS voices — do not impersonate; mark synthetic speech where it matters; keep the CC BY attribution

Original Source

Conclusion

If you need offline Tamil TTS that fits phone / screen-reader / voice-agent budgets and you’re fine with literary studio voices plus CC BY credit, Kuyil is today’s open drop — start with kuyil-20m, then A/B the 5M and 1.4M cards on your device.

—Anabel ♡