Chinese Accents Identification with Modified MFCC

摘要

It is well known that performance of Chinese speech recognition system fluctuates sharply with variation of accents. Hence it is feasible to identify accents before recognition. Based on our previous analysis [1], it is discovered that the first two formants are more sensitive to Chinese accents than others. A modified Mel-frequency cepstral coefficients (MFCC) algorithm is proposed by increasing the filter distribution in lower and middle frequency range to accommodate the sensitivity of Chinese accents. Comparing with the GMM system based on traditional MFCC, the error rate of the GMM system based on modified MFCC declines by 1.8%.