WaveGlow

WaveGlow is a flow-based generative model for speech synthesis developed by NVIDIA. Introduced in 2018, it functions as a neural vocoder that converts mel-spectrograms into time-domain audio waveforms. The model combines ideas from Glow, a flow-based framework using invertible 1x1 convolutions and affine coupling layers, with concepts from WaveNet to model complex audio distributions. WaveGlow enables high-quality speech synthesis without autoregressive sampling during generation.

Technically, WaveGlow uses a sequence of invertible 1x1 convolutions to permute data channels and a stack of affine coupling layers that transform a simple Gaussian latent variable into an audio waveform conditioned on a mel-spectrogram. The overall transformation is invertible, and training maximizes the log-likelihood of the observed waveform given the conditioning. This architecture allows parallelizable sampling: after training, one can draw a latent vector from a standard normal distribution and apply the inverse transformations to produce audio, typically faster than autoregressive vocoders.

In a text-to-speech pipeline, WaveGlow serves as the vocoder component that converts predicted mel-spectrograms into waveform. It is often used in conjunction with a separate acoustic model that generates mel-spectrograms from text, such as a Tacotron-style system. NVIDIA released open-source implementations to facilitate research and practical deployment, highlighting the model’s balance of audio quality and generation speed on GPU hardware.

Limitations include substantial computational and memory requirements for training, as well as potential artifacts or noise that can appear in some setups. Quality and stability depend on data quality, conditioning features, and careful tuning of training hyperparameters.