Speech synthesis that ditches autoregression for normalizing flows

What it does

WaveGlow is a PyTorch implementation of a flow-based generative network that turns mel-spectrograms into raw audio. Instead of predicting each audio sample conditioned on all previous samples—the way WaveNet does—it uses a reversible flow model to generate the waveform in parallel. The authors claim it matches the best publicly available WaveNet implementation in mean opinion score while synthesizing audio at 1200 kHz on an NVIDIA V100.

The interesting bit

The project marries OpenAI’s Glow with DeepMind’s WaveNet: it keeps the high-quality inductive biases of WaveNet’s architecture but drops the autoregressive training and inference bottleneck. By using only a single network and a single cost function—maximizing training data likelihood—the authors argue the training pipeline stays simple and stable, which is rarer in generative audio than one might hope.

Key highlights

• Synthesizes speech from mel-spectrograms without autoregression
• Claims 1200 kHz sample generation on an NVIDIA V100 GPU
• Mean Opinion Scores reportedly on par with the best public WaveNet implementation
• Supports mixed-precision (fp16_run) and multi-GPU training via NCCL
• Published pre-trained model and mel-spectrogram datasets available

Caveats

• Multi-GPU training is only tested on single-node NCCL setups
• Older checkpoints require conversion via convert_model.py to work with the current fused residual and skip connections
• Depends on NVIDIA Apex, so portability is tied to the CUDA/PyTorch ecosystem

Verdict

Researchers and TTS builders who need WaveNet-level quality without the serial inference penalty should look here; if you are not on NVIDIA hardware or need proven multi-node scaling, this is less compelling.