WaveGrad: Estimating Gradients for Waveform Generation [pdf]

Nanxin Chen, Yu Zhang, Heiga Zen, Ron J. Weiss, Mohammad Norouzi, William Chan, ICLR 2021

Abstract: This paper introduces WaveGrad, a conditional model for waveform generation through estimating gradients of the data density. This model is built on the prior work on score matching and diffusion probabilistic models. It starts from Gaussian white noise and iteratively refines the signal via a gradient-based sampler conditioned on the mel-spectrogram. WaveGrad is non-autoregressive, and requires only a constant number of generation steps during inference. It can use as few as 6 iterations to generate high fidelity audio samples. WaveGrad is simple to train, and implicitly optimizes for the weighted variational lower-bound of the log-likelihood. Empirical experiments reveal WaveGrad to generate high fidelity audio samples matching a strong likelihood-based autoregressive baseline with less sequential operations.

Illustration of WaveGrad's waveform generation in only 6 refinement iterations:

Note: To obtain the best audio quality, listen with headphones. Consider reducing the volume for the first few iterations below as they are mostly white noise.

Text: Here are the match lineups for the Colombia Haiti match.

n=0

              

n=1

              

n=2

              
n=3               
n=4               
n=5               
n=6               

Illustration of WaveGrad's waveform generation in 50 refinement iterations:


Samples from WaveGrad conditioned on a continuous scalar indicative of the noise level:

Note: Only a single model is trained. Different rows correspond to different iterative refinement schedules for inference.
Samples from both ground truth features (top) and predicted features (bottom) are provided.

Text Weekends at twenty three fifty. Here are the match lineups for the
Colombia Haiti match.
On Friday night in Bridgeport expect a temperature of minus four degrees Fahrenheit.
Reference
WaveGrad Base (Linear 1000)


WaveGrad Base (Linear 50):


WaveGrad Base (Fib 25):


WaveGrad Base (Manual 6):


Samples from WaveGrad conditioned on a discrete iteration index:

Note: Each row corresponds to an individual model that is trained with a particular iterative refinement scheudle in mind.
Samples from both ground truth features (top) and predicted features (bottom) are provided.

Text Weekends at twenty three fifty. Here are the match lineups for the
Colombia Haiti match.
On Friday night in Bridgeport expect a temperature of minus four degrees Fahrenheit.
Reference
WaveGrad Base (Linear 1000)


WaveGrad Base (Linear 50):


WaveGrad Base (Fib 25):


Samples from other non-autoregressive baselines

Note: Samples from both ground truth features (top) and predicted features (bottom) are provided.

Text Weekends at twenty three fifty. Here are the match lineups for the
Colombia Haiti match.
On Friday night in Bridgeport expect a temperature of minus four degrees Fahrenheit.
Multi-band MelGAN:


MelGAN:


Parallel WaveGAN:


Samples from WaveGlow/WaveFlow

Note: Due to the time limitation, we trained WaveGlow for 300k steps and WaveFlow for 1.8 million steps on 4 NVIDIA V100 GPUs using official implementation. Samples from both ground truth features (top) and predicted features (bottom) are provided.

Text Weekends at twenty three fifty. Here are the match lineups for the
Colombia Haiti match.
On Friday night in Bridgeport expect a temperature of minus four degrees Fahrenheit.
WaveGlow:


WaveFlow: