Multi-attention feature fusion speaker recognition method

Abstract

The invention provides a multi-attention feature fusion speaker recognition method, which comprises the following steps of: constructing a deep speaker characterization model which comprises a feature extractor and a speaker classifier, taking Fbank features as input of the deep speaker characterization model, extracting the Fbank features as speaker characterization through the feature extractor, in a training stage, using a speaker classifier to map the speaker representation to a speaker tag, and constructing a loss function to optimize a deep speaker representation model; and in a test stage, comparing similarity between speaker representations by adopting a cosine distance, and judging whether the speakers are the same speakers or not according to a threshold value. According to the method, weighted fusion is performed on the features of different branches through multi-attention feature fusion, and the multi-attention feature fusion comprises a space attention mechanism and a channel attention mechanism, so that effective information in each branch is enhanced, and speaker recognition performance with higher robustness is obtained.

Description

technical field [0001] The invention belongs to the technical field of artificial intelligence, and in particular relates to a speaker recognition method for fusion of multiple attention features. Background technique [0002] With the development of voice technology, more and more devices support voice control, such as smartphones, smart speakers and smart cars. In order to improve the security of voice control, speaker recognition technology is often added to these smart devices as a front-end service to ensure that only specific speakers use these voice services. Speaker recognition is a very hot topic, and many methods have been proposed to solve this problem. [0003] The core step of speaker recognition is to extract speaker representations from speech signals. In the early days, the probability density function of the speech signal was used to describe the identity information of the speaker, and the Gaussian Mixture Model-Universal Background Model (GMM-UBM) was on...

AI Technical Summary

Problems solved by technology

At present, researchers have found that multi-branch features can help the model learn more discriminative speaker representations. This method maps features to different branches through different convolution kernel parameters, and each branch is processed separately. Finally, the The features of each branch are fused. When the multi-branch features are fused, the traditional method adopts the method of adding or splicing, which cannot give full play to the characteristics of the multi-branch features.

Images