Estimating Gradients for Waveform Generation [pdf]


[Submitted on 2 Sep 2020]

Download PDF

Abstract: This paper introduces WaveGrad, a conditional mannequin for waveform generation
thru estimating gradients of the knowledge density. This mannequin is constructed on the
prior work on procure matching and diffusion probabilistic models. It begins from
Gaussian white noise and iteratively refines the signal through a gradient-essentially based fully mostly
sampler conditioned on the mel-spectrogram. WaveGrad is non-autoregressive, and
requires finest a fixed series of generation steps at some level of inference. It might per chance well per chance
use as few as 6 iterations to generate excessive fidelity audio samples. WaveGrad is
straightforward to coach, and implicitly optimizes for the weighted variational
lower-traipse of the log-likelihood. Empirical experiments show WaveGrad to
generate excessive fidelity audio samples matching a stable likelihood-essentially based fully mostly
autoregressive baseline with much less sequential operations.

Submission history

From: Nanxin Chen [view email]

Wed, 2 Sep 2020 17: 44: 10 UTC (598 KB)

Read More

Leave A Reply

Your email address will not be published.