A dynamic processing approach to phoneme recognition (part I)--Feature extraction

This paper presents a new approach to phonemic feature extraction, which is one of the essential components of the study to develop an automatic speech recognition system. The feature extraction method is composed of three levels, i.e., frequency analysis, peak-band enhancement, and local peak extraction. The latter two levels consist of a multistage processing structure and are made by using analogical formulations to mechanical dynamics. Extracted features include local peak frequencies and their amplitudes, and local valley frequencies and their amplitudes, which are estimated by the computational process with no feedback loop and no backtracking. The feasibility is first confirmed by using synthesized speech, then by application to continuous vowels, multisyllable words, voiceless stop consonants, and voiceless fricatives. The advantages are discussed in comparisons with other methods.