PrasadTTS
Indian-language speech generated in your browser. The model downloads once and runs on your CPU — nothing is sent to a server.
Downloaded once and kept in your browser cache — later visits load it from disk.
stage: —
Telugu is my own model — 110M parameters, distilled and quantised to 4-bit. Hindi and Tamil are still training. The other languages are Kyutai's released models. Each has a single fixed voice.
Prasad Vittaldev
Credits
- pocket-tts by Kyutai (MIT) — the architecture, the Mimi codec, the training and distillation recipes, and the English, French, German, Italian, Portuguese and Spanish models offered above.
- pocket-tts-onnx-export by KevinAHM (MIT) — the ONNX export scripts every graph here was produced with.
- Hindi model: saryps-labs/pocket-tts-hindi — a Hindi student of pocket-tts trained on AI4Bharat IndicVoices. All the Hindi speech quality is theirs; only the int4 packaging is mine.
- ONNX Runtime Web — runs the model in your browser.
- Telugu speech data: IndicTTS (SPRING Lab, IIT Madras) and SYSPIN (IISc Bangalore).
- My models: pocket-tts-telugu-female-syspin, plus the int4 ONNX packs every language above is downloaded from — the full collection.