基于二阶隐马尔可夫模型的清浊音恢复算法

Voiced/Unvoiced Parameters Recovery Based on Second-Order Hidden Markov Model

为了解决低速率语音编码中比特受限的问题,提出了一种基于二阶隐马尔可夫模型的清浊音参数恢复算法。该算法采用二阶隐马尔可夫模型,通过归一化的能量参数和LPC倒谱系数估计出序列中的全带清浊音判决和各个子带的清浊音度。解码器实现该算法后,编码器就无需对清浊音参数进行量化传输,从而节约了比特数。实验结果表明,该算法比基于GMM模型的算法能更好地恢复出清浊音信息,全带清浊音误判率减少了5%～20%,合成语音的MOS分比用5 bit的矢量量化(VQ)算法提高了0.03左右,达到了在节约比特数的同时也提高了语音质量的效果。

In order to solve the problem of limited number of bits in low bit rate speech coding, an algorithm using second - order Hidden Markov ModeI（HMM2） to recover the voiced/unvoiced parameters is proposed in this paper. The algorithm uses the normalized energy and linear prediction coding（LPC） coefficients to estimate the full-band V/U classification and the sub-band BPVC value. The algorithm can be implemented in the decoder, saving the bits originally used by V/U parameters and reducing the bit rate of speech coding. Experimental re- suits show that the algorithm proposed can reduce the V/U classification error rate by 5 % - 20 % compared with the GMM algorithm, and improve the mean opinion score（MOS） of the synthesized speech signal by about 0.03 compared with the 5bit vector quantization（VQ）, thereby greatly improves the estimation performance of the V/ U parameters.