Comparative study of several distortion measures for speech recognition

In this study we compared several different spectral distortion measures including the Itakura-Saito (IS), the log likelihood ratio (LLR), the likelihood ratio (LR), the cepstral (CEP), and two perceptually based distortion measures, the weighted likelihood ratio (WLR) and the weighted slope metric (WSM) distortion measures, in terms of their effects on the performance of a standard dynamic time warping (DTW) based, isolated word, speech recognizer. Two modifications of the basic forms of each measure were also investigated, namely a Bark-scale frequency warping and the incorporation of surasegmental energy information. All distortion measures and their modifications were tested on an alpha-digit vocabulary, 4-talker, telephone recording data base. The results can be summarized as: (1) All LPC-based distortion measures performed reasonably well. The LLR and WSM distortion measures gave the highest recognition accuracy, while the IS distortion measure gave the lowest score; (2) Whereas the addition of suprasegmental energy information helped the recognition performance, the use of gain and absolute loudness degraded the performance; (3) Bark-scale frequency warping did not perform as well as its unwarped counterpart; (4) The WLR distortion measure did not perform as well as its unweighted counterpart. L Introduction Since it was first ietroduced, the Itakura-Saito distortion measure 1] has played a key role in speech coding, analysis, synthesis and recognition. Several studies were conducted to investigate the relationship between different LPC-based distortion measures and to study their propertis from a theoretical point of view 12,31. It is the goal of this research to compare several basic distortion measures (including two recently proposed, perceptually based measures [4,51) and to study their influence on the performance of an isolated word, DTW based, speech recognition system. We also tested two modifications of the basic distortion measures: Bark-scale frequency warping of the LPC-derived distortion measure, and incorporation of suprasegmental energy information. II. Spectral Distortion Measures 2.1 Itakura-Saito Distortion Measure The maximum likelihood distortion measure, also known as the Itakura-Saito distortion measure, was first used for short-time spectral estimation of speech signals. The measure, denoted as dis, is: dis(S,f)f [-+lnj_l] e where Si,, (A) is the short-time spectral density (or periodograni) of an input speech signal, and f (A) a a2 -Ii + a,e" + + ae"l2 — Al2 is the spectral density function of a corresponding pth-order all-pole model. Defining d as the log spectral distance between S,, (A) and f(T), at frequency A, i.e. 2.2 The Log Likelihood Ratio (LLR) and the Likelihood Ratio (LR) Distortion Measures TRa1 dLLR Ifln d15(f,f3f') = in a