Fast Indic TTS — front-end ➜ back-end demo
A code-mixed text normalization + grapheme-to-phoneme front-end (LoRA-fine-tuned Sarvam-1) chained to a fast Grad-TTS acoustic back-end. Code & write-up: github.com/Abhijit89Kumar/fast-indic-tts.
The NFE slider sets the number of diffusion steps: Grad-TTS defaults to 50, but ~16 retains most of the quality at roughly 3× lower latency. Try the 2nd-order Heun sampler too. Note: the acoustic model is English-only (LJSpeech); the normalization and phonemes are the code-mixed contribution, and the English voice will approximate Hindi words.
4 50
Examples
| Input text (English or code-mixed Hinglish) | Sampler | NFE (diffusion steps) |
|---|