This repository provides an implementation of TacotronV2 and WaveRNN for Chinese text-to-speech synthesis, targeting researchers and developers working with Mandarin speech. It enables the generation of high-quality, natural-sounding Chinese speech from text, with features for speaker adaptation and improved long-sentence modeling.

How It Works

The system combines TacotronV2 for converting Chinese text to mel-spectrograms and WaveRNN for generating audio waveforms from these mel-spectrograms. It preprocesses Chinese text by converting characters to Pinyin, handling multi-syllable words and numbers. For improved long-sentence synthesis, it explores alternative attention mechanisms like Gaussian Mixture Attention and Forward Attention. Speaker adaptation is achieved by fine-tuning TacotronV2 on new data.

Quick Start & Requirements

• Install dependencies via requirements.txt.
• Run synthesis with python tacotron_synthesize.py --text '...'.
• Requires Python and TensorFlow (tested with 1.14.0) or PyTorch.
• Preprocessing and training scripts are provided for both TacotronV2 and WaveRNN.

Highlighted Details

• Implements TacotronV2 with PyTorch and TensorFlow.
• Offers speaker adaptive training for TacotronV2.
• Explores alternative attention mechanisms (GMM, Discretized Graves, Forward Attention) to improve long-sentence modeling.
• Supports mixed Chinese and Pinyin input.
• Includes TensorFlow Serving + Flask for deployment.

Maintenance & Community

• Last updated October 2020.
• No explicit community links (Discord/Slack) or active maintenance signals are present in the README.

Licensing & Compatibility

• The README does not explicitly state a license.
• Compatibility for commercial use or closed-source linking is not specified.

Limitations & Caveats

The project was last updated in October 2020, suggesting potential compatibility issues with newer library versions. The README mentions that Location-sensitive attention in TacotronV2 has limitations with long sentences, though alternative attention mechanisms are explored.