M.-T. Lin (Taiwan)
INITIAL Recognition, Segmental Probability Model(SPM), Pattern Comparison.
For Mandarin INITIAL (Consonant) recognition, a two-stage algorithm with multiple features was proposed. We classified Mandarin INITIAL into seven classes, each with similar phonetic and acoustic characteristics. In the first stage of the proposed algorithm, INITIAL classification was conducted by computing features to construct subscore for each feature and for each INITIAL class, and then we defined for the decision making of INITIAL classification. Our multi-speaker experiments showed that a high correct percentage (98.8% for the top two selections) could be achieved in our first-stage INITIAL classification. In the second stage, with the selected top two INITIAL classes as candidates, detailed INITIAL recognition is performed by feature extraction of cepstral coefficients followed by the pattern comparison using the continuous hidden Markov model (CHMM) or the segmental probability model (SPM). The experimental results showed that both SPM and CHMM can achieve comparable accuracy percentage of 90.6%, but the SPM requires only about 1/8 processing time. Furthermore, the proposed two-stage algorithm is superior to the one-stage algorithm (omitting the first stage) by about 3% accuracy rate increasing and 12% processing time saving.
Important Links:
Go Back
