A Novel Voiceprint Verification Technology Through Deep Neural Network

This paper presents an integrated neural network for voiceprint verification. The system implements two types of deep architecture: ResNet-18 and SincNet to extract the acoustic features. The triplet loss function is used to distinguish same-speaker and different-speaker pairs based on cosine similarity during the training phase. Experiments on three different datasets reveal that integrated system exceeds the baseline of DNN-based i-vector. The system reduces equal error rates (EERs) from the baseline method by 59.7, 54.5 and 58% on the datasets—Voxceleb1, LibriSpeech and AISHELL-1, respectively. In addition, the integrated model decreases the EERs in the single models.