论文部分内容阅读
作者讨论了人工神经网络矢量量化技术在多带激励语音压缩编码算法中的实际应用。采用Ｋｏｈｏｎｅｎ自组织特征映射神经网络技术对语音参数中的谱包络参数进行量化，利用Ｋｏｈｏｎｅｎ自组织特征映射神经网络具有的聚类特性，提出一种初始码本抽取和码本训练的实际算法，训练出具有明显拓扑结构的码本。利用语音的帧间相关性和训练网络的结构特性，提出一种称为“邻域搜索法”的快速码字搜索算法。实验表明，这种矢量量化算法使码本搜索时间下降为ＬＢＧ全搜索的１／４～１／６，量化质量有所提高。
The author discusses the practical application of artificial neural network vector quantization in multi-band excitation speech compression coding algorithm. Kohonen self-organizing feature mapping neural network is used to quantify spectral envelope parameters in speech parameters. By using Kohonen’s self-organizing feature mapping neural network with clustering characteristics, an initial algorithm of codebook extraction and codebook training is proposed. Train a codebook with a clear topology. Using the inter-frame correlation of speech and the structural characteristics of training networks, a fast codeword search algorithm called “neighborhood search method” is proposed. Experiments show that this vector quantization algorithm reduces the codebook search time to 1/4 ~ 1/6 of the full search of LBG, and improves the quantization quality.