How ai4bharat/indic-tts works
AI4Bharat Indic-TTS provides training and inference for Indian-language speech synthesis, with published models for 13 languages. The documented workflow formats and analyzes the IndicTTS dataset, configures FastPitch acoustic-model and HiFi-GAN vocoder training, then serves text-to-speech through an API or Triton. The inference engine normalizes and transliterates text, segments it, synthesizes audio, and postprocesses the result. The graph also preserves documented preprocessing, training, and deployment capabilities; unsampled training and socket-proxy implementation details are represented only where the tree supports a path. Edges cite the README or shown caller code.
Main entry points: inference.py TTS engine · server.py HTTP API · configs Training configs · text.py Text normalization
Components 22
Dataset preparation
- IndicTTS dataset
- Dataset formatter
preprocessing/FormatDataset-IndicTTS.ipynb - Dataset analysis
preprocessing/AnalyzeDataset-IndicTTS.ipynb
Model training
- FastPitch training
main.py - HiFi-GAN training
vocoder.py - Training configs
configs
Speech inference
- TTS engine
inference/src/inference.py - Request models
inference/src/models/request.py - Audio responses
inference/src/models/response.py - Text normalization
inference/src/utils/text.py - Text translation
inference/src/utils/translator.py - Paragraph splitting
inference/src/utils/paragraph_handler.py - Audio postprocessing
inference/src/postprocessor/postprocessor.py - Voice activity detection
inference/src/postprocessor/vad.py - Audio denoiser
inference/src/postprocessor/denoiser.py
Serving interfaces
- HTTP API
inference/server.py - Inference sample
inference/sample.py - Triton model
inference/triton_server/triton_repo/tts/1/model.py - Triton client
inference/triton_server/client.py - Socket proxy
inference/socket_proxy/te_endpoint.py - Browser client
inference/socket_proxy/front_end/main.js
Other components
- TTS caller
Connections 19
- IndicTTS dataset → Dataset formatter: formats
README.md - IndicTTS dataset → Dataset analysis: analyzes
README.md - Training configs → FastPitch training: configures
README.md - Training configs → HiFi-GAN training: configures
README.md - TTS caller → HTTP API: sends request
inference/server.py - HTTP API → TTS engine: calls inference
inference/server.py - HTTP API → Request models: accepts
inference/server.py - TTS engine → Request models: reads
inference/src/inference.py - TTS engine → Audio responses: constructs
inference/src/inference.py - TTS engine → Text normalization: normalizes
inference/src/inference.py - Text normalization → Text translation: translates text
inference/src/utils/text.py - TTS engine → Paragraph splitting: splits text
inference/src/inference.py - TTS engine → Audio postprocessing: processes audio
inference/src/inference.py - Audio postprocessing → Voice activity detection: trims silence
inference/src/postprocessor/postprocessor.py - TTS engine → Audio denoiser: denoises audio
inference/src/inference.py - TTS caller → Inference sample: runs example
- TTS caller → Triton client: uses client
- Triton model → TTS engine: calls inference
inference/triton_server/triton_repo/tts/1/model.py - Browser client → Socket proxy: connects