Throat Microphone Signals for Syllable Recognition Using Linear Prediction Cepstrum

The performance of standard Automatic Speech Recognition (ASR) systems using the Normal Microphone (NM) degrades even if the ambience is slightly noisy. In contrast to the NM speech, the Throat Microphone (TM) speech is unaffected by such an ambience. This paper explores the feasibility of using the TM speech for developing robust ASR systems in these conditions. This ASR system may also be useful in situations where NM cannot be used. A hidden Markov model based system is developed to recognize the 145 basic syllabic units of the Indian language Hindi. Recognition systems are also developed to categorize the syllabic units into broad categories, based on the vowel, place of articulation, and manner of articulation of the syllables. The performance of the TM based system is found to be comparable to the performance of the NM based system.