能量参数解码端HMM估计算法

HMM estimation of energy contours in speech decoders

在低速率语音编码算法中,如何对特征参数进行有效的量化表示是影响声码器合成语音质量的关键因素。该文提出一种能量参数解码端恢复算法,它利用线谱频率(linespectral frequency,LSF)和清浊音判决参数(unvoiced/voiced decision,U/V)估计能量参数的变化轨迹。该算法利用特征参数之间的相关性,采用隐Markov模型(hiddenMarkov model,HMM)描述LSF、U/V和能量参数之间的统计特性,通过对能量进行解码端恢复,省去量化所需的比特数,从而提高特征参数的整体量化性能。测试结果表明:能量参数解码端恢复算法能够将150b/s混合激励线性预测编码算法(mixed excitation linear prediction,MELP)的合成语音平均意见得分(mean opinion score,MOS)提高0.042。该算法应用于超低速率声码器参数量化是可行的。

Low bit rate speech coding must effectively quantize the parameters.This article presents an energy contour estimation algorithm to predict changes of speech energy from the line spectral frequency（LSF） and the unvoiced/voiced（U/V） decision parameters.The statistical properties of the energy,the LSF and the U/V decision parameters are characterized based on the hidden Markov model（HMM） which uses the correlations between different parameters.The algorithm properly estimates the energy contour,which contributes to quantization of the decoder parameters.Tests show that the energy contour estimation algorithm improves the mean opinion score（MOS） of the synthesized speech for the mixed excitation linear prediction（MELP） vocoder at a 150 b/s bit rate by 0.042,which shows that this algorithm improves parameter quantization in ultra low bit rate vocoders.