Simulating Speech Coders Using Neural Networks

Abstract

Speech coding algorithms are developed and optimised to satisfy many applications' specific requirements. By using these requirements to analyse different types of speech coders operat ing at various bit rates, the most efficient speech coding scheme for a particular applica tion can be selected. Analysing speech coders is a time-consuming, expensive process because many of the tests are based on human percep tion of the reconstructed speech. An efficient, cost-effective method of performing this analy sis is to use a neural network. After a neural network is trained using characteristic param eters of the reconstructed speech signal, it can be used to classify a speech coder's performance in terms of its application requirements. This greatly simplifies the process of selecting a speech coder for a particular application. This paper shows how a neural network can be used to identify the speech quality and signal-to- noise ratio of a speech waveform produced by a speech coder, and how it can be used to deter mine if a speech coder is suitable for a particu lar application.