Speech Feature Extraction Based on Linear Prediction Residual

摘要

Linear prediction coding (LPC) is the core technology in speech processing, which has been successfully applied in speech recognition, synthesis and coding. LPC coefficient can well represent the speaker's vocal tract information, and is widely used in the field of speaker recognition. However, the epiphytic LPC residual is often ignored. This paper shows that the LPC residual contains information, which can reflect the characteristics of the speaker himself. We extracted new feature parameters (Second moment, third moment) from the LPC residual, combined with LPC coefficients, and input them into a speaker recognition system based on the GRU network. Cross entropy loss is used as a loss function to train the classifier. The experimental results show that the combined parameters effectively improve the system recognition rate by 5% relative to the LPC coefficients.