Speaker recognition method under linear transformation of identity vector x-vector

Abstract

The invention relates to a speaker recognition method under linear transformation of an identity vector x-vector. The method comprises the main steps that feature extraction is carried out on speech,and the identity vector x-vector and identity vector i-vector of the speech are extracted respectively; parallel factor analyzer training is carried out by using the x-vector and i-vector of a same speaker; a parameter corresponding to the x-vector in a parallel factor analyzer is selected, and on the basis of the parameter, linear transformation is carried out on the identity vector x-vector to obtain an xl-vector; the new identity vector xl-vector is trained to obtain a PLD model; feature extraction and x-vector extraction are carried out on the to-be-tested speech, the identity vector x-vector is input into a linear transducer obtained in the training stage to obtain the new identity vector xl-vector, and finally, the identity vector xl-vector is input into the PLD model obtained in thetraining stage, thereby obtaining a final result. The speaker recognition method under the linear transformation of the identity vector x-vector has the advantages of improving the recognition performance of speaker recognition while ensuring that the memory requirements and computing speed are similar to those of a baseline system.

Description

technical field [0001] The present invention relates to speaker identification technology in biometric identification, and more specifically relates to a speaker identification technology under linear transformation of identity vector x-vector. Background technique [0002] Voice is the most direct and convenient way for human beings to communicate. It has attracted the attention of various research institutions for its unique convenience, economy, accuracy and other advantages. The research on speech signal processing is of great significance to the promotion of human-computer interaction and the development of artificial intelligence. For this reason, the related fields of speech signal processing, such as speech recognition, speech coding, speech synthesis, speaker recognition, etc., have received more and more attention and theoretical research. Speaker recognition, also known as voiceprint recognition, aims to authenticate the identity of each speaker based on the uniq...

AI Technical Summary

Problems solved by technology

Different papers have studied how to improve the system performance under the x-vector model. Studies have shown that superposition of i-vector and x-vector models or fusion of PLDA scores can improve system performance. However, this method is designed for two systems and requires a lot of The memory requirements, while the calculation speed will also be affected

Subsequently, more research improved the robustness of x-vector through data expansion, but this method is affected by the recognition environment

Images