DeepVoice3 is an open-source text-to-speech project on GitHub that implements fully convolutional neural network speech synthesis models in PyTorch. Maintained by Ryuichi Yamamoto (r9y9), the repository has earned roughly 2,000 stars and is one of the best-known community implementations of Baidu’s Deep Voice 3 architecture for end-to-end speech synthesis.
The project provides a complete training and inference pipeline for convolutional sequence-to-sequence TTS models. Instead of recurrent networks, the architecture relies entirely on convolutional layers for both the text encoder and the mel-spectrogram decoder, which makes training significantly faster and parallelizable on modern GPUs. The codebase supports both single-speaker and multi-speaker setups, and ships with presets and preprocessing scripts for popular speech corpora. A project page with generated audio samples is linked from the repository, so visitors can listen to output quality before cloning anything. The repository was last updated in late 2023 and remains a widely referenced baseline for neural TTS research.
The repository also includes recipes for preparing datasets, extracting features and monitoring training through alignment and spectrogram visualizations, which makes it easier to diagnose problems while a model learns to speak clearly.
DeepVoice3 is completely free and open source. Anyone can clone the repository from GitHub, train models on their own data and use the generated voices without paying license fees or subscribing to a hosted service.
DeepVoice3 suits speech researchers, ML engineers and hobbyists who want a proven, hackable PyTorch codebase for training custom text-to-speech voices, whether for academic experiments, prototyping voice products, or learning how convolutional TTS architectures work in practice.