Introduction

Feel that pocket open across more than English and Mandarin? Dheeyantra Research Labs just pushed Pravega-TTS (dheeyantra/dhee-pravega-tts)—a CosyVoice 3 fine-tune aimed at 11 Indian languages plus Arabic and Indian English, without throwing away the European languages the base already spoke.

Zero-shot all the way: a voice isn’t a fixed speaker ID. It’s a short reference clip plus that clip’s exact transcript. Output lands at 24 kHz mono.

What’s Inside Pravega-TTS

Base: FunAudioLLM/Fun-CosyVoice3-0.5B-2512 (Apache-2.0). Fine-tuned: the text-to-token LM (llm.pt) and flow-matching decoder (flow.pt). Vocoder, speech tokenizer, speaker encoder, and text encoder stay stock.

Languages targeted: Assamese, Bengali, English, Gujarati, Hindi, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil, Telugu, Arabic—plus German, Spanish, French, Russian carried from the base.

Their indicative round-trip CER (Whisper-large-v3 after synthesis) shows real Indic gains where baselines existed—mean over seven languages with a CosyVoice 3 baseline: 0.940 → 0.382. Regression on base languages stayed mild (0.030 → 0.038). They flag small n, Whisper’s weak Telugu/Malayalam/Assamese scoring, and Odia as unscorable by Whisper—not necessarily broken audio.

Licence nuance matters: Dhee Research-Only Licence v1.0 (not commercial) because some Indic training data ties to IIT-Madras Indic TTS non-commercial terms. Base CosyVoice bits remain Apache-2.0. Consent-only cloning—don’t impersonate.

Pro Tip

Keep the fine-tune prefix: “You are a helpful assistant.<|endofprompt|>” then the reference clip’s transcript verbatim. Mismatched transcripts are the usual reason a clone falls apart. Clean 4–12 s mono clips win.

How to Get Started with Pravega-TTS

  1. Clone CosyVoice with submodules and install requirements.
  2. snapshot_download("dheeyantra/dhee-pravega-tts") and load via CosyVoice3.
  3. Pick a voice from voice_bank.jsonl (six redistributable Assamese/Odia/Punjabi clips ship) or bring your own consented reference.
  4. Run inference_zero_shot and save 24 kHz WAV.

Model: dheeyantra/dhee-pravega-tts · Base: Fun-CosyVoice3 · Runtime: CosyVoice GitHub

Conclusion

Pravega-TTS is the open CosyVoice 3 fork that finally treats Indic languages like first-class pocket, not an afterthought—zero-shot cloning, honest CER caveats, research-only licence. Pull the weights, bring a consented clip, and let that unfiltered motion speak.

—Anabel ♡