Reproduce “WaveGlow: A Flow-based Generative Network for Speech Synthesis” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Ryan Prenger, Rafael Valle, Bryan Catanzaro

NVIDIA researchers developed WaveGlow, a neural vocoder combining normalizing flows with WaveNet-style architectures to generate high-quality audio. This model produces audio at 520 kHz on an NVIDIA V100 GPU, over 25 times faster than real-time, while achieving audio quality comparable to WaveNet with a simpler single-network, likelihood-maximization training approach.

Key insights

  • Normalizing flow models can effectively learn invertible mappings between simple distributions and complex audio waveforms, enabling both exact likelihood computation during training and parallel sample generation.
  • Combining a WaveNet-style subnetwork within affine coupling layers allows for high-quality audio generation while mel-spectrogram conditioning guides the synthesis process.
  • A single maximum likelihood objective function can train a generative vocoder to produce high-fidelity audio at speeds significantly faster than real-time, simplifying previous complex training setups.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.