基于非负矩阵分解的2kb/s波形内插语音编码算法

2kb/s Waveform Interpolation Speech Coding Based on Non-negative Matrix Factorization

在波形内插(Waveform Interpolation,WI)语音编码器中,如何低延时、高精度并且低复杂度的分解和量化特征波形(Characteristic Waveform,CW)一直是该编码模型的研究热点和难点.本文提出用非负矩阵分解(Non-negative MatrixFactorization,NMF)方法来分解语音特征波形.该分解方法仅需要当前帧的语音信号,不会给编码器带来额外的延时;为了提高分解精度,本文在CW分解之前先对CW按照其子帧的最大基音周期进行分类,然后按不同类别进行分解.另外,本文结合耳蜗模型提出了NMF的基矢量分带初始化算法,将CW的分解精度提高到与二阶奇异值分解相当的水平;为了降低WI编码器的计算复杂度,本文去除了传统WI编码器中的特征波形对齐模块,同时将NMF的分解阶数设定为16以折中CW分解的计算复杂度和分解精度.最后,本文基于矩阵量化技术,对非负矩阵分解后的编码矩阵采用分裂式矩阵量化方案来量化.主观A/B测试表明,本文提出的2kb/s NMF-WI编码器的合成语音质量接近于2.4kb/s SVD-WI编码器.MOS分测试表明,本文提出的2kb/s NMF-WI编码器的合成语音质量稍差于2.4kb/s MELP编码器.

In WI coding scheme, how to decompose and quantize the characteristic wavefonns with low delay, low complexity and high precision have always been a hot research topic. The characteristic waveform decomposition based on non-negative matrix factorization is proposed in this paper. This CW decomposition method doesn＇t bring any additional delay to WI coder;In order to improve decomposition precision, the CW is firstly classified according to the maximum pitch of its sub-frames before being decomposed. Besides, band-partitioning initialization constraints are set to basis vectors before NMF is carried out, and this has made the CW decomposition precision of NMF-based method be comparable with that of 2 ranks of SVD; In order to reduce the computational complexity of WI coder, the CW alignment procedure is removed in our NMF-WI coder, and the factorization rank of NMF is set to 16 as a trade-off between computational complexity and decomposition precision.In the end,the low dimensional coding matrix is quantized by splitting matrix quantization scheme. The subjective A/B listening tests show that the proposed 2kb/s NMF-WI coder can give smooth speech with quality close to 2.4kb/s SVD-based WI coder.Mean Opinion Score test results indicate that the performance of proposed coder is a little worse than that of 2.4kbps MELP coder.