WaveGrad 2: Iterative Refinement for Text-to-Speech Synthesis

[jinglescode]

Paper

Link: https://arxiv.org/pdf/2106.09660.pdf
Year: 2021

Summary

• text-to-speech synthesis, synthesizes the waveform directly without using hand-designed intermediate features (e.g., spectrograms)

Methods

3 modules

• encoder: sequence input, extracts representations
• resampling: match input to output
• decoder: generate waveform

encoder:

• 3 conv + batchnorm + dropout
• LSTM
• zoneout regularization

resampling

• Gaussian upsampling introduced in the non-attentive Tacotron

decoder

• consist upsampling blocks and downsampling blocks

Results

• tradeoff between fidelity and speed by varying the number of refinement steps
• experiments demonstrate that WaveGrad 2 is capable of generating high fidelity audio, comparable to strong baselines