Speaker-Independent Consonant Classification and Recognition for Mandarin Syllables

摘要

In this paper, a two-stage scheme for consonant classification and recognition has been developed, which can be used as a subsystem for Mandarin syllable recognition of speaker-independent by acoustic phonetic approach. The first stage discriminate and assign the total 21 Mandarin consonants into seven classes. There are seven kinds of manners of articulation for Mandarin consonants, namely aspirate plosive, unaspirate plosive, aspirate affricate, unaspirate affricate, unvoiced fricative, voiced fricative and sonorant. The averaged power, duration, zero-crossing rate, average magnitude difference function, and energy ratio between low and high frequency bands were selected as the acoustic features for our consonant classification algorithm. Further discrimination in the second stage enables the identification of all Mandarin consonants using continuous density hidden Markov models (CHMM) and segmental probability model (SPM) alternatively. Experimental result showed that this approach could provide significant improvements in recognition performance and speed. This is achieved by considering the special characteristcs of the target vocabulary.