Research on Speaker Recognition Based on LPCMCC and PSRCC Fusion Features

Mel frequency cepstral coefficient (MFCC) is an important feature in speech speaker recognition, and many people like to use MFCC and neural network algorithms to form speaker recognition systems. However, the MFCC coefficient loses the phase information of the speech signal during the extraction process. Therefore, this article proposes a speech recognition method that combines LPC Mel frequency cepstral coefficients (LPCMCC) and phase based spectral root cepstral coefficients (PSRCC). After combining the two, the amplitude and phase information of the speech signal can be compatible and used as input data for the neural network. This article introduces an attention mechanism based on ResNeXT to adjust the weights of feature parameters, highlight key areas, and improve recognition performance. Finally, the AISHELL dataset was used for training and validation, and the experimental results showed that the proposed combined features and improved ResNeXT can effectively improve the accuracy of speaker recognition.