Competitive Learning Methods for Efficient Vector Quantizations in a Speech Recognition Environment

摘要

This paper presents a comparison of three competitive learning methods for vector quantizations of speech data in an efficient way. The analyzed algorithms were two batch methods (the Lloyd LBG algorithm and the Neural Gas method) and one on-line technique (K-means algorithm). These methods obtain reduced subsets of codewords for representing bigger data sets. The experiments were designed for speaker dependent and independent tests and consisted in evaluating the reduced training files for speech recognition purposes. In all the studied cases, the results shown a reduction of learning patterns of near 2 orders of magnitude respect to the original training sets without heavily affecting the speech recognition accuracy. The savings in time after using these quantization techniques, made us to consider this reduction results as excellent since they help to approximate the speech matching responses to almost real time. The main contribution of this work refers to an original use of competitive learning techniques for efficient vector quantization of speech data and so, for reducing the memory size and computational costs of a speech recognizer.