基于深度学习的多声源并行化声纹辨别方法

Multi-source parallelized voiceprint identification method based on deep learning

随着大数据时代的到来,快速而有效地辨别声纹已经成为智能感知领域的重要需求,而传统神经网络和单拾音器系统的辨别精度不高,样本数据量大,其运算速度严重制约了系统的实时性.文中方法通过拾音阵列获取目标声源的位置和时频域信息,利用GPU并行构造掩蔽函数,实现信号数据级融合,强化目标语音特征,然后将多通道的MFCC(mel-frequency cepstral coefficient)声纹参数进行特征级融合,输入深度信念网络(deep belief network,DBN)进行训练和识别,同时使用CUDA(compute unified device architecture)平台对DBN的训练过程进行了并行优化.该方法能在多声源环境下全面地提取目标声纹,有效提高声纹辨别准确率,缩短数据训练耗时,保证了系统实时性.该方法为大数据环境下语音信号高性能处理提供了一种实现方式.

With the coming of the era of big data,the distinguishing of voiceprint in a quick and effective way has become the important demand in intelligent perception field.However,the identification accuracy of traditional neural network and single pick-up system is not high enough,and the sample data size is large;therefore,the calculation speed has seriously restrained the timeliness of the system.By obtaining the location and time-frequency domain information of the targeted sound source by pick-up array,this method utilizes GPU parallel structure sheltered function to realize the data level integration of signal,to strengthen the characteristics of targeted voice,to conduct feature-level fusion to multipath MFCC voiceprint parameters,and to input deep belief network(DBN)for training and identification,and utilizes CUDA platform to conduct parallel optimization to the training process of DBN.This method can abstract targeted voiceprint in multiple sound source environment,to improve the accuracy of voiceprint identification,and shorten data training time-consuming,to ensure the timeliness of system.This method provides an effective method for the high performance treatment of voice signal in big data environment.