ಕನ್ನಡ Text-to-Speech
Kokoro-82M · Kannada

A fast, small, open TTS voice for Kannada — 82M parameters, runs on CPU, trained for under $100 of compute. Prosody comes from a trained diffusion sampler, not a fixed style vector.

Code & full write-up Model weights The journey (all findings)

🔊 Listen

Generated by the released model — 82M parameters, single pass, no post-processing or cherry-editing.

ಇಂದು ಬೆಳಿಗ್ಗೆ ನಾನು ಕಾಫಿ ಕುಡಿಯುತ್ತಾ ಪತ್ರಿಕೆ ಓದಿದೆ. ಹೊರಗೆ ಮಳೆ ಬರುತ್ತಿತ್ತು, ಮತ್ತು ವಾತಾವರಣ ತುಂಬಾ ಹಿತವಾಗಿತ್ತು.

“This morning I read the newspaper over coffee. It was raining outside, and the weather was lovely.”

ಬೆಂಗಳೂರಿನಲ್ಲಿ ಇಂದು ವಾತಾವರಣ ತುಂಬಾ ಚೆನ್ನಾಗಿದೆ.

“The weather in Bengaluru is very nice today.”

ನೀವು ಊಟ ಮಾಡಿದ್ದೀರಾ?

“Have you eaten?”

Why it matters

Model size	82M parameters — runs on a phone-class CPU
Speed	RTF 0.05–0.4 on GPU · a few seconds/sentence on CPU
Intelligibility	ASR round-trip CER 0.00 on clean text — a Kannada ASR transcribes the synthesis back exactly
Total training compute	< $100 (Colab + small cloud-GPU jobs)
Data	~11 h public Kannada speech (IndicTTS + AI4Bharat Rasa), 17 speakers
License	Code Apache-2.0 · weights inherit training-data terms

Kannada is language #1 of a planned open family: one small model · one voice · every major Indian language, running locally — no cloud API required.

Run it yourself

# mac: brew install espeak-ng   ·   debian: apt install espeak-ng
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 namaskara.wav --best-of 3