A Method for Speaker Recognition Based on the ResNeXt Network Under Challenging Acoustic Conditions

Speaker recognition is an indispensable technology for biometrics. It distinguishes individuals based on their vocal patterns. In this paper, a joint confirmation method based on the Akaike Information Criterion (AIC) of reconstruction error (REE) and time complexity (AIC-Time joint confirmation method) is proposed to select the optimal decomposition rank of NMF. Furthermore, non-negative matrix factorization (NMF) is applied to the spectrogram to generate speaker features. The network for speaker recognition is based on Convolutional Neural Networks combining Squeeze Excitation (SE) blocks with ResNeXt, and the best combination is explored experimentally. The SE block conducts a channel-level adaptive adjustment of the feature maps, reducing redundancy and noise interference while improving feature extraction efficiency and accuracy. The ResNeXt convolutional neural network concurrently executes multiple convolutional kernels, acquiring richer feature information. The experimental results demonstrate that compared to speaker recognition based on Gaussian mixture models (GMM), Visual Geometry Group Network (VGGNet), ResNet, and SE-ResNeXt using spectrograms, this method increases the accuracy by an average of 5.8% and 16.24% under the overlaid of babble and factory1 noise with different signal-to-noise ratios, respectively. In the short speech test, the test set is short speech of 1s and 2s, and the noise is superimposed. Compared with other methods, the recognition rate is increased by an average of 8.67% and 11.72%, respectively.