基于矢量量化的说话人识别系统硬件实现  

Hardware implementation of speaker recognition system based on vector quantization

作　　者：何赞园 王凯 吉立新 HE Zanyuan;WANG Kai;JI Lixin(Strategic Support Force Information Engineering University of PLA,Zhengzhou 450002,China)

机构地区：[1]中国人民解放军战略支援部队信息工程大学,河南郑州450002

出　　处：《现代电子技术》2022年第1期171-175,共5页Modern Electronics Technique

基　　金：国家自然科学基金资助项目(61803384);国家自然科学基金资助项目(61801515)。

摘　　要：在实际应用中,矢量量化取决于码本设计、码字搜索和码字索引分配等三个关键技术,其中码本设计和码字搜索最为关键。说话人识别是一种典型的多码本应用场景,并且说话人识别算法相对比较规则,因此可采用全搜索算法进行运算,从硬件设计的角度讲,全搜索算法虽然运算量大,但算法只涉及乘、加和比较操作,采用硬件实现是切实可行的。首先对现有矢量量化算法的硬件复杂度进行分析,针对全搜索算法实现了矢量量化硬件设计,并结合该方法实现了多路电话信道中说话人的实时判别。利用FPGA对所提方法进行实验验证,结果表明所提矢量量化的硬件解决方案相比传统方案可有效提升矢量量化速度与效率。In practical application,vector quantization depends on codebook design,codeword search and codeword index allocation,among which codebook design and codeword search are the most critical. Speaker recognition is a typical multi codebook application scenario. Because the speaker recognition algorithm is relatively regular,the full search algorithm can be used for operation. In the perspective of hardware design,the full search algorithm only involves multiplication,addition and comparison operations although it needs a large amount of computation,so the hardware implement is feasible. The hardware complexity of the existing vector quantization algorithm is analyzed. The hardware design of vector quantization is realized according to the full search algorithm. The real-time identification of the speaker in multi-channel telephone channel is realized in combination with the hardware implementation based on vector quantization. The FPGA(field-programmable gate array) is used to verify the proposed method. The experimental results show that the proposed hardware solution of vector quantization can effectively improve the speed and efficiency of vector quantization in comparison with the traditional solutions.

关 键 词：说话人识别 矢量量化 LBG算法 码本设计 码字搜索 全搜索算法 FIFO FPGA 

分 类 号：TN912-34[电子电信—通信与信息系统]