Low-Bitrate Speech Compression with a Glottal Pulse Autoencoder

Author(s):

Michael S. Lee

Abstract:

In this report, a convolutional neural network with a custom voicing layer is used to compress English speech to a rate of 1500 bits per second. The model consists of an autoencoder that converts speech input into a quantized latent space and then decodes with voice-like glottal sound and noise layers multiplied by a formant mask. The input and output of the model are magnitude short-time Fourier transform spectrograms. Each 32-ms frame of the input is approximately mapped to an element of the bottleneck layer and quantized to 48 bits via 4 additive layers of 12-bit vector codebooks.