DATA COMPRESSION TECHNIQUES FOR ISOLATED AND CONNECTED WORD RECOGNITION (SPEECH SIGNALS, TEMPLATE MATCHING)

Dynamic Time Warping or Dynamic Programming (DP) is one of the most efficient techniques for automatic speech recognition. Most commercial speech recognition systems in use today are based on this technique. The main problem associated with the practical application of this technique is the limited size of the vocabulary. This problem arises from the large memory space required for reference storage as well as the massive number of computations (multiplications and additions) needed to calculate the DP equations.

A major part of this thesis is concerned with solving the above problem for both Isolated and Connected Word Recognition systems. The techniques proposed to overcome these problems are based on either: (i) reducing or compressing speech data that results from a linear prediction coefficient (LPC) analysis or (ii) using another speech processing technique that produces feature vectors for the speech frames of smaller dimensions than those produced by known analysis techniques. The principal contributions to the subject area reported in this thesis are compounded in the following techniques: (1) A segmentation technique for isolated word recognition systems; (2) A data compression technique for connected word recognition systems; and (3) A speech processing technique that results in feature vectors of lower dimensions than those produced by an LPC or filter bank analysis.

In addition to the main contributions listed above, methods for solving some of the problems associated with dynamic programming and end point detection for both isolated and connected word recognition are described.

The work concentrates on the use of template matching schemes for speech recognition systems. Full details of the different components for template matching systems are introduced. The systems are based on using LPC analysis and DP for the matching process. The mathematical and computational techniques associated with LPC analysis are discussed and a detailed explanation for isolated word DP is given. Different DP equations are tested and compared. In the case of connected word recognition, a one-path DP algorithm is used. (Abstract shortened by UMI.)