Residual speech signal compression: an experiment in the practical application of neural network technology

Abstract

Neural networks are a popular area of research today. However, neural network algorithms have only recently proven valuable to application problems. This paper seeks to aid in the process of transferring neural network technology from research to a development environment by describing our experience in applying this technology. The application studied here is Speaker Identity Verification (SIV), which is the task of verifying a speaker's identity by comparing the speaker's voice pattern to a stored template. In this paper, we describe the application of the back-propagation neural network algorithm to one aspect of the SIV problem, called Residual Compression (RC). The RC problem is to extract useful features from a part of the speech signal that was not utilized by previous SIV systems. Here, we describe a neural network architecture, pre-processing algorithm, training methodology, and empirical results for this problem. We also present a few guidelines for the use of neural networks in applied settings.