benjaminwan/ChineseTtsTflite

Implements a pipeline combining FastSpeech2 and Tacotron2 models with MelGAN vocoder for real-time Chinese speech synthesis, built with Kotlin and Jetpack Compose on TensorFlow Lite. Supports offline inference with selectable model variants—FastSpeech2 optimized for mid-range devices with faster inference, while Tacotron2 prioritizes audio naturalness for higher-end hardware. Includes aggressive binary optimization, reducing TFLite dependencies from 109.6MB to 14.8MB through selective op compilation.

393 stars. No commits in the last 6 months.