Understanding WaveGrad

WaveGrad is a conditional model for waveform generation which estimates gradients of the data density. It is built on the concept of diffusion probabilistic models and score matching.

Key Concepts

1. Diffusion Models: WaveGrad defines a forward process that adds noise to the data and a reverse process that denoises it.
2. Gradient Estimation: The core idea is to estimate the gradient of the log-density of the data distribution.

Comparison with WaveNet

Unlike WaveNet, which is autoregressive, WaveGrad is non-autoregressive, allowing for faster inference in some settings, although it requires multiple iterations (steps) to generate high-quality audio.

Conclusion

WaveGrad represents a significant step forward in high-fidelity audio synthesis, offering a trade-off between inference speed and sample quality through the number of refinement steps.