WaveGlow is NVIDIA’s open-source neural vocoder: a flow-based generative network that turns mel spectrograms into natural-sounding speech audio. Released under the BSD-3-Clause license with more than 2,300 GitHub stars, the repository was last updated in 2023 and remains a reference implementation in the text-to-speech literature.
In a classic TTS pipeline, one model predicts a mel spectrogram from text and a second model โ the vocoder โ converts that spectrogram into a waveform. WaveGlow combines ideas from Glow-style normalizing flows and WaveNet into a single network trained with a simple maximum-likelihood objective, avoiding the adversarial training instability of GAN vocoders and the slow autoregressive sampling of WaveNet. The result generates audio faster than real time on modern GPUs while matching the quality of the best public vocoders of its generation, and NVIDIA published pretrained models trained on the LJSpeech dataset alongside the code.
WaveGlow is free and open source under the BSD-3-Clause license.
WaveGlow is for speech researchers and TTS engineers who need a proven, well-cited vocoder baseline, and for students studying how normalizing flows apply to audio generation.
NVIDIA’s release includes a pretrained model trained on the LJSpeech corpus at a 22.05 kHz sampling rate, alongside training code that pairs WaveGlow with a Tacotron 2 acoustic model for a complete text-to-speech system. The model has roughly 87 million parameters and became a standard component in NVIDIA’s NeMo and Jarvis speech pipelines. Researchers frequently cite WaveGlow as the baseline that later parallel vocoders โ such as HiFi-GAN and DiffWave โ set out to beat.