Signal Representation Attribute Extraction and the Use Distinctive Features for Phonetic Classification

Signal Representation Attribute Extraction and the Use Distinctive Features for Phonetic Classification 1 Helen M. Meng, Victor W. Zue, and Hong C. Leung of Spoken Language Systems Group Laboratory for Computer Science Massachusetts Institute of Technology Cambridge, Massachusetts 02139 ABSTRACT The study reported in this paper addresses three issues related to phonetic classification: 1) whether it is important to choose an appropriate signal representation, 2) whether there are any ad- vantages in extracting acoustic attributes over directly using the spectral information, and 3) whether it is advantageous to intro- duce an intermediate set of linguistic units, i.e. distinctive fea- tures. To restrict the scope of our study, we focused on 16 vowels in American English, and investigated classification performance using an artificial neural network with nearly 22,000 vowels tokens from 550 speakers excised from the TIMIT corpus. Our results indicate that 1) the combined outputs of Seneff's auditory model outperforms five other representations with both undegraded and noisy speech, 2) acoustic attributes give similar performance to raw spectral information, but at potentially considerable com- putational savings, and 3) the distinctive feature representation gives similar performance to direct vowel classification, but po- tentially offers a more flexible mechanism for describing context dependency. INTRODUCTION The overall goal of our study is to explore the use of dis- tinctive features for automatic speech recognition. Distinc- tive features are a set of properties that linguists use to clas- sify phonemes [1,13]. More precisely, a feature is a minimal unit which distinguishes two maximally-close phonemes; for example/b/and/p/are distinguished by the feature [voicE]. Sounds are more often confused in relation to the number of features they share, and it is believed that around 15 to 20 distinctive features are sufficient to account for phonemes in all languages of the world. Moreover~ the values of these fea- tures, such as [+HIGH] or [-aOUND], correspond directly to contextual variability and coarticulatory phenomena, and of- ten manifest themselves as well-defined acoustic correlates in the speech signal [3]. The compactness and descriptive power of distinctive features may enable us to describe contextual influence more parsimoniously, and thus to make more effec- tive use of available training data. 1This research was supported by DARPA under Contract N00014- 82-K-0727, monitored through the Office of Naval Research. In order to fully assess the utility of this linguistically well-motivated set of units, several important issues must be addressed. First, is there a particular spectral representa- tion that is preferred over others? Second, should we use the spectral representation directly for phoneme/feature classi- fication, or should we instead extract and use acoustic cor- relates? Finally, does the introduction of an intermediate feature-based representation between the signal and the lexi- con offer performance advantages? We have chosen to answer these questions by performing a set of phoneme classification experiments in which conditional variables are systematically varied. The usefulness of one condition over another is in- ferred fl'om the performance of the classifier. In this paper, we will report our study on the three ques- tions that …

Content truncated to 500 words