Kokoro-82M · Kannada (ಕನ್ನಡ)
Fast, small, open Kannada TTS: Kokoro-82M (distilled StyleTTS2) fine-tuned for Kannada with a trained diffusion prosody sampler — the component stock Kokoro drops, and the difference between robotic and natural prosody.
- 🔊 Samples: https://huggingface.co/spaces/Sammy-Mohan/kannada-tts
- 📖 Code, recipe & the full journey (all findings): https://github.com/sammy4321/Kokoro-Indic-Fine-Tuning
Files
| file | purpose |
|---|---|
kokoro_kannada_00005_inference.pth |
model weights (optimizer stripped, inference-ready) |
config_kn.yml |
StyleTTS2 architecture/config |
ref_spk9.wav |
the style-reference clip = the shipped voice (full-bandwidth speaker — important, see repo) |
Usage
Use the inference script (handles the fork's architecture code + diffusion sampling + best-of-N):
git clone https://github.com/sammy4321/Kokoro-Indic-Fine-Tuning
cd Kokoro-Indic-Fine-Tuning && pip install -r demo/requirements.txt huggingface_hub
python inference/synthesize.py --text "ನಮಸ್ಕಾರ!" --out out.wav --best-of 3
Details
82M params · 24 kHz · trained on ~11 h public Kannada speech (SPRINGLab IndicTTS + AI4Bharat Rasa, 17 speakers) for <$100 of compute. ASR round-trip CER 0.00 on clean text. Known gaps: numerals-as-digits and heavy English code-switching need a text-normalization frontend (roadmap).
License: code Apache-2.0; weights are derived from the datasets above — check their terms before commercial use; attribution required. Built on StyleTTS2 (MIT), Kokoro-82M (Apache-2.0), and the kikiri-tts recipe.