基于大津算法和深度学习的开集声纹识别自适应阈值计算方法

针对开集声纹识别的自适应阈值计算问题, 提出一种基于大津算法和深度学习的阈值计算方法. 首先, 采用三层受限Boltzmann机堆叠而成的深度置信网络从Mel倒谱系数中提取语音深层特征; 其次, 通过Gauss混合模型计算特征的相似度值, 对特征的相似度值使用大津算法计算阈值. 实验结果表明, 该方法识别效果较理想, 与等错误率计算阈值方法相比, 具有更高的识别准确率.

Calculation Method of Adaptive Threshold for Open Set Voiceprint Recognition Based on Otsu Algorithm

Aiming at the problem of adaptive threshold calculation for open set voiceprint recognition, we proposed a threshold calculation method based on Otsu algorithm and deep learning. Firstly, deep features of speech were extracted from Mel cepstrum coefficients by using deep confidence network stacked by three layers of restricted Boltzmann machines. Secondly, the similarity value of features was calculated by Gaussian mixture model, and Otsu algorithm was used to calculate the threshold value of the similarity value of the features. The experimental results show that the proposed method has a better recognition effect, compared with the method of calculating the threshold value with equal error rate, the method has higher recognition accuracy.