IndicVoice-82M

IndicVoice

A decoder-only neural TTS model for Indian languages, built on the Kokoro-82M architecture with a native Indic G2P frontend.

Author: Kushal Kant Bind — Chandigarh University Project: GSoC 2026, Sugar Labs GitHub: Bindkushal/indic-voice

---

Quick Start

pip install git+https://github.com/Bindkushal/indic-g2p.git
pip install git+https://github.com/Bindkushal/indic-voice.git
apt-get install espeak-ng

from indicvoice import IndicPipeline
pipeline = IndicPipeline(lang_code="hi", repo_id="Bindkushal/IndicVoice-82M")
for gs, ps, audio in pipeline("नमस्ते दुनिया", voice="af_heart"):
    import soundfile as sf
    sf.write("output.wav", audio, 24000)

---

Files in This Repo

---

Supported Languages

---

Architecture

• Base: Kokoro-82M (StyleTTS2 + ISTFTNet), Apache 2.0
• G2P: indic-g2p — native Indic phonemizer
• Fallback: espeak-ng for OOV words
• Parameters: 82M
• Sample rate: 24000 Hz

---

Citation

@misc{bind2026indicvoice,
  title={IndicVoice: Decoder-Only Neural TTS with Native G2P for Indian Languages},
  author={Kushal Kant Bind},
  year={2026},
  institution={Chandigarh University},
  note={GSoC 2026, Sugar Labs}
}

Acknowledgements

• hexgrad/kokoro — base TTS architecture (Apache 2.0)
• AI4Bharat — IndicVoices-R dataset
• IIT Madras — IndicTTS dataset

License

Apache 2.0