关键词: 重音检测, Mel频率倒谱系数（MFCC）, 相关谱感知线性预测（RASTA-PLP）, 短时谱特征

Abstract: Accent is a critically important component of spoken communication, and plays a very important role in spoken communication. In order to verify the effect of short-time spectrum feature set based on auditory model in Chinese accent detection method, this paper uses MFCC (Mel frequency cepstrum coefficient) algorithm and RASTA-
PLP (relative spectra perceptual linear prediction) algorithm to extract each voice segment of short-time spectrum information, and builds short-time spectrum feature sets based on MFCC algorithm and RASTA-PLP algorithm. Then, it chooses NaiveBayes classifier to model the two feature sets, and chooses the classes with maximum a posteriori probability as the object’s class. This classification method makes full use of the related phonetic features of speech segment. Short-time spectrum feature set based on MFCC and short-time spectrum feature set based on RASTA-
PLP respectively achieve 82.1% and 80.8% accent detection accuracy on ASCCD (annotated speech corpus of Chinese discourse). The experimental results indicate that short-time spectrum features based on MFCC and short-time spectrum features based on RASTA - PLP can be used for Chinese accent detection research.

Key words: accent detection, Mel frequency cepstrum coefficient (MFCC), relative spectra perceptual linear prediction (RASTA-PLP), short-time spectrum features