Cory S. Myers

Performance tradeoffs in dynamic time warping algorithms for isolated word recognition

The technique of dynamic programming for the time registration of a reference and a test pattern has found widespread use in the area of isolated word recognition. Recently, a number of variations on the basic time warping algorithm have been proposed by Sakoe and Chiba, and Rabiner, Rosenberg, and Levinson. These algorithms all assume that the test input is the time pattern of a feature vector from an isolated word whose endpoints are known (at least approximately). The major differences in the methods are the global path constraints (i.e., the region of possible warping paths), the local continuity constraints on the path, and the distance weighting and normalization used to give the overall minimum distance. The purpose of this investigation is to study the effects of such variations on the performance of different dynamic time warping algorithms for a realistic speech database. The performance measures that were used include: speed of operation, memory requirements, and recognition accuracy. The results show that both axis orientation and relative length of the reference and the test patterns are important factors in recognition accuracy. Our results suggest a new approach to dynamic time warping for isolated words in which both the reference and test patterns are linearly warped to a fixed length, and then a simplified dynamic time warping algorithm is used to handle the nonlinear component of the time alignment. Results with this new algorithm show performance comparable to or better than that of all other dynamic time warping algorithms that were studied.

A level building dynamic time warping algorithm for connected word recognition

Dynamic time warping has been shown to be an effective method of handling variations in the time scale of polysyllabic words spoken in isolation. This class of techniques has recently been applied to connected word recognition with high degrees of success. In this paper a level building technique is proposed for optimally time aligning a sequence of connected words with a sequence of isolated word reference patterns. The resulting algorithm, which has been found to be a special case of an algorithm previously described by Bahl and Jelinek, is shown to be significantly more efficient than the one recently proposed by Sakoe for connected word recognition, while maintaining the same accuracy in estimating the best possible matching string. An analysis of the level building method shows that it can be obtained as a modification to the Sakoe method by reversing the order of minimizations in the two-pass technique with some subsequent processing. This level building algorithm has a number of implementation parameters that can be used to control the efficiency of the method, as well as its accuracy. The nature of these parameters is discussed in this paper. In a companion paper we discuss the application of this level building time warping method to a connected digit recognition problem.

An investigation of the use of dynamic time warping for word spotting and connected speech recognition

Several variations on algorithms for dynamic time warping have been proposed for speech processing applications. In this paper two general algorithms that have been proposed for word spotting and connected word recognition are studied. These algorithms are called the fixed range method and the local minimum method. The characteristics and properties of these algorithms are discussed. It is shown that, in several simple performance evaluations, the local minimum method performed considerably better then the fixed range method. Explanations of this behavior are given and an optimized method of applying the local minimum algorithm to word spotting and connected word recognition is described.

Modeling chaotic systems with hidden Markov models

The problem of modeling chaotic nonlinear dynamical systems using hidden Markov models is considered. A hidden Markov model for a class of chaotic systems is developed from noise-free observations of the output of that system. A combination of vector quantization and the Baum-Welch algorithm is used for training. The importance of this combined iterative approach is demonstrated. The model is then used for signal separation and signal detection problems. The difference between maximum likelihood signal estimation and maximum a posteriori signal estimation using a hidden Markov model is illustrated for a nonlinear dynamical system.<<ETX>>

Speaker independent connected word recognition using a syntax-directed dynamic programming procedure

A method for speaker independent connected word recognition is described. Speaker independence is achieved by clustering isolated word utterances of a 100 speaker population. Connected word recognition is based on a syntax-directed dynamic programming algorithm which matches the isolated word templates to sentence length utterances. The method has been tested on an artificial task-oriented language based on a 127 word vocabulary. Four subjects, two men and two women, spoke a total of 209 sentences comprising 1750 words. At an average speaking rate of 171 words/min over dialed-up telephone lines, a correct word recognition rate of 97 percent was observed.

Connected word recognition using a level building dynamic time warping algorithm

The technique of dynamic time warping has proven itself reliable and robust for a wide variety of isolated word recognition tasks. Recently extensions of the algorithm have been investigated for application to the problem of connected word recognition. In this paper a level building technique is proposed for optimally aligning a test pattern, consisting of a sequence of connected words, with a sequence of isolated word reference patterns. This algorithm is shown to be significantly more efficient than the one proposed by Sakoe while solving the exact same problem. Implementation parameters for the level building algorithm are presented and the effectiveness of the proposed algorithm for connected digit recognition is experimentally verified.

Knowledge-based pitch detection

Knowledge-based signal processing combines the symbolic manipulation facilities of artificial intelligence with the numerical methods of signal processing. This paper considers pitch detection as a problem for developing techniques for signal/symbol interaction. Specifically an overview of the design of the Pitch Detectors Assistant program is discussed, along with some strategies followed in the program. for the use of information derived from both signals and symbols.

Knowledge based speech analysis and enhancement

This paper describes a system for speech analysis and enhancement which combines signal processing and symbolic processing in a closely coupled manner. The system takes as input both a noisy speech signal and a symbolic description of the speech signal. The system attempts to reconstruct the original speech waveform using symbolic processing to help model the signal and to guide reconstruction. The system uses various signal processing algorithms for parameter estimation and reconstruction.

An automated directory listing retrieval system based on recognition of connected letter strings

In this paper we describe a system which is capable of recognizing spoken spelled names from a directory of names. These names are spelled in a connected fashion, i.e., without any pause between letters, over a dialed‐up telephone line. This system uses a level building dynamic time warping algorithm to perform a time registration between the input speech and a sequence of letters drawn from isolated word templates. The initial result of the matching procedure is a string of letter classes. Letter classes are formed by breaking the alphabet up into groups of acoustically similar letters. The directory of names is also sorted by letter classes. After the input string is classified into letter classes, a second pass is performed to determine the optimally matching name. This system was tested on four talkers each of whom spoke a given set of 50 names. The names were spoken at both a deliberate and at a normal talking rate. Recognition was performed in both a speaker‐trained and a speaker‐independent manner....

Performance trade‐offs in dynamic time warping algorithms for isolated word recognition

The technique of dynamic programming for time registration of a reference and a test utterance has found widespread use in the area of discrete word recognition. Recently a number of variations on the basic time warping algorithms have been proposed by Sakoe and Chiba, and Rabiner, Rosenberg, and Levinson. These algorithms all assume the test input is an isolated word whose endpoints are known (at least approximately). The major difference in the methods are the global path constraints (i.e., the region of possible paths), the local continuity constraints on the path, and the distance weighting and normalization used to give the overall minimum distance. The purpose of this investigation is to study the effects of such variations on the performance of different algorithms for a realistic speech data base. The performance index is based on speed of operation, memory requirements, and recognition accuracy of the algorithm. Preliminary results indicate, in most cases, only small differences in performance among the various methods.