一种基于小波变换和隐Markov模型的声调识别方法

A TONE RECOGNIZER USING WAVELET TRANSFORM AND HIDDEN MARKOV MODEL

本文给出了一种基于小波变换和隐Markov模型(HMM)的声调识别方法。根据小波变换检测信号突变的性质,充分利用多分辨率分析,准确可靠地实现了基音检测;采用分划Gauss混合(PGM)概率密度函数的HMM进行汉语声调识别,推导出用PGM函数的Viterbi算法的简化递推式。在匹配计算量大大减小的情况下,特定人的四声识别率为97.22％,非特定人达到94.47％。

This paper presents a tone recognizer for Mandarin speech using a combination of wavelet transforming and hiddefi Markov modeling techniques. The evaluation of pitch periods is exactly performed by a pitch detector which is based on the singularity detection of signal and multiresolution analysis with wavelet transform. The hidden Markov models with partitioned Gaussian mixtures(PGM) are used for tone recognition. In implementing the Viterbi algorithm for HMM's, a recursive relation is derived to improve the computation efficiency, where the accuracy is 97.22%, 94.47% for speaker-dependent and speaker-independent tone recognition respectively.