摘要
本文介绍了一种基于矢量量化技术的说话人识别系统,在特征参数上,选用动、静态倒谱互相补偿,充分描述说话人声道模型,同时选用能描述说话人声带激励源特点的基音参数,以全面反映说话人特征。识别时,对三个参数进行优化组合,构成最佳的归一化联合失真进行判决.实验证明,对16人进行测试,当语音长度为2秒时,识别率高达96.8%.
In this paper, a vector quantization based speaker is proposed.On the selectionof the, we chose the instantaneous and transitional cepstral representations which can give complemenetaryinformation of vocal tract, and chose the pitch period which can describe the property of vocal cords. Usingthere parameters, the speaker's charateristics can be more sufficiently provied. With opitimized combination ofthree features,the best normalized combination distance is used to make a recognition decision. with 16-speakers'database, about 96.8% identification rate is achieved for 2 seconds test speech.
出处
《信号处理》
CSCD
2000年第1期85-89,共5页
Journal of Signal Processing