Speech Signal Compression Using Neural Networks

Abstract

We propose a speech compression technique based on vector quantization. A neural network with unsupervised learning is used to implement the vector quantizer. Some general issues concerning the vector quantization problem are presented, as well as some basic aspects related to speech signal processing. The idea of using a codebook to perform speech compression is introduced, and the use of a 2-dimensional self-organizing Kohonen map to generate the codebook is proposed. Finally, simulation results are presented, giving some insights on the best network initialization and training strategies, as well as the best network topology for this problem.