WaveRNN is a widely used open-source PyTorch implementation of DeepMind’s WaveRNN neural vocoder, with 2,100+ stars on GitHub, bundled with a Tacotron implementation to form a complete text-to-speech system you can train yourself.
Neural vocoders turn mel spectrograms into listenable waveforms, and WaveRNN — from DeepMind’s Efficient Neural Audio Synthesis paper — was designed to do it fast enough for practical deployment without giving up audio quality. This repository is the community implementation that most open-source TTS projects learned from: a PyTorch WaveRNN vocoder plus a Tacotron front end that predicts spectrograms from text, so the full pipeline from sentence to spoken audio lives in one place. A quick-start script synthesizes speech from a text file in minutes using pretrained pieces, while the full training workflow — preprocessing the LJSpeech dataset, training Tacotron, generating a GTA dataset, then training the vocoder — is documented step by step with training visualizations. Last updated in 2022, it is best read as a stable, complete reference implementation for learning or building on classic neural TTS architecture.
WaveRNN is free and open source under the MIT license. You train and run it on your own GPU; there is no hosted service, paid tier or usage fee for the code.
Speech synthesis researchers, TTS hobbyists and audio ML engineers who want a complete, trainable classic neural TTS system — or a vocoder reference implementation to build on.