indic-diffusion-tts

Diffusion-based text-to-speech for Indic languages.

Layout

.
├── front-end-normalization/       text normalisation, G2P, tokenisation
└── back-end-diffusion/            acoustic model (Grad-TTS-style) + vocoder
    ├── dsp-playground/            Tasks 1 & 2 — STFT, mel-spec, Griffin-Lim
    └── grad-tts-baseline/         Upstream Grad-TTS + HiFi-GAN + latency benchmark

Status

Component	Status
DSP playground (mel + phase)	done
Grad-TTS baseline (local, 50 steps)	done — 1.98 s / 5.65 s of audio, RTF 0.35x on RTX 3050
Rule-based front-end "Before" gallery	done — see front-end-normalization/FAILURE_REPORT.md
LLM-based front-end ("After")	todo
Triton-fused diffusion decoder	todo
Vocoder swap (HiFi-GAN → BigVGAN?)	todo

See:

• back-end-diffusion/dsp-playground/ — spectrogram pipeline notebook, math PDF, Griffin-Lim phase-problem demo.
• back-end-diffusion/grad-tts-baseline/ — runnable Grad-TTS inference and end-to-end latency log.
• front-end-normalization/ — naive rule-based pipeline (AI4Bharat indic-numtowords) and the five sentences that break it.