Residual error and batch normalization-based neural network noisy vocal print identification method

Abstract

The invention discloses a neural network noisy vocal print recognition method based on residual error and batch normalization, which comprises the following steps: preparing a data set in a noisy environment, and respectively extracting corresponding Mel-frequency cepstral coefficients from voice signals; constructing a ResTDNN network structure: after the construction is completed, training the network by using a natural gradient random descent method, after the training is completed, capturing vocal print information of a speaker by using voice information about 10 seconds through x-vector extracted by dimension reduction, and adopting a cross entropy loss function; and performing dimension reduction by using linear discriminant analysis, performing length normalization on an x-vector after the dimension reduction, training linear probability analysis, respectively calculating likelihood functions of two voices from different spaces, and evaluating a recognition system by adopting equal error rate and minimum detection cost. According to the method, the robustness of the time delay neural network in the noise environment can be improved, the neural network degeneration and gradient disappearance phenomenon requirements are relieved, and the vocal print recognition rate is improved.

Description

technical field [0001] The invention belongs to the field of voiceprint recognition, and in particular relates to a neural network noise pattern recognition method based on residual error and batch normalization. Background technique [0002] Most of the traditional voiceprint recognition technologies are based on the identity factor (Identity Vector, I-Vector), but the modeling ability of this method needs to be optimized. In recent years, the use of deep neural network (DNN) to capture the speaker's speech features is a big boom, but this method increases the computational complexity while meeting the training requirements. Using time-delayed neural network (Time Delayed Neural Network, TDNN) embedding, and using a deep neural network to train the speaker recognizer to extract speaker information can effectively improve DNN. The performance of the TDNN algorithm is stronger than that of I-Vector in all aspects, but when there is strong noise interference in the recognitio...

AI Technical Summary

Problems solved by technology

The performance of the TDNN algorithm is stronger than that of I-Vector in all aspects, but when there is strong noise interference in the recognition environment, the robustness effect is not good, and as the network depth increases, the network is prone to gradient disappearance and degradation.

Images