基于动态阈值失量量化的说话人识别

Vector quantization based on the dynamic threshold of speaker recognition

在基于矢量量化的说话识别系统所选用的LBG算法中,码本分裂时的阈值是影响初始码本生成的重要因素之一,而传统方式所采用的阈值不容易确定,且需要进行大量的实验来获得经验值。提出在一定范围内动态地,随机地产生阈值的方法来改进初始码本形成策略,并结合差分倒谱参数建立说话人识别模型。实验结果表明该方法在识别率得到一定改善的前提下,训练时间及识别时间有了明显改善。

Code splitting threshold is one of the important factors to initialize codebook in Speaker Recognition based on the Vector Quantitation （ VQ）, but traditional threshold is not easy to determine and needs a large number of experiments to determine the value. This paper used dynamic and random method to select the threshold in a certain range, and combined with differential cepstrum thresholds to establish speaker recognition model. The results show that given the method improves the recognition rate, the training time and the recognition time have improved significantly.