说话人识别和确认是信号处理中研究的热点之一,但有关文献表明识别效率并不是很高,而且训练和识别的语音要求都比较长,距离实际应用还有一定差距.分析了说话人识别中有关参数的选取对识别结果的影响,采用线性预测倒谱和基音参数共同作为识别参数,并采用矢量量化,改进了线性预测倒谱距离的加权函数,提供了与文本无关的说话人识别系统.最后给出了实验结果和有关分析,在低噪声时识别正确率可达99%以上,在高噪声时也能达到98%以上的正确率.%Speaker recognition and identification is one of the research hot topics in signal processing. But the related documents indicate that its recognising efficiency has limitations, and long speech is required for training and recognition, there is still certain distance apart from the practical application. In this article we analyse the influence of selecting relevant parameters in speaker recognition on the outcome of recognition, and provide a speaker recognition system independent to the text which uses linear prediction (LP) cepstrum and pitch parameter as the joint recognition parameters, and quantises vectors by the vector quantization (VQ), improves the weighting function of LP eepstrum distance. The experimental results and relevant analysis are given in the last part of the paper. In low noise environment the recognition correct rate approaches 99％ or higher, and that is also higher than 98％ in condition of high-noise.
展开▼
