Predictive Vector Quantized Variational AutoEncoder for Spectral Envelope Quantization

Abstract

The Predictive Vector Quantized Variational AutoEncoder is proposed to improve the reconstruction error of the conventional VQ-VAE. The proposed model can predict the current data from the previous data. The performance of the quantized spectral envelope parameters of the high-quality 48 kHz WORLD vocoder is evaluated. The results indicate that the Predictive Vector Quantized Variational AutoEncoder has a lower distortion with four target bitrates in term of log-spectral distortion, compared with the conventional VQ-VAE.