Audio feature extraction method and device based on re-parameterized decoupling mode

Abstract

The invention discloses an audio feature extraction method and device based on a re-parameterized decoupling mode. The method comprises the following steps: acquiring a to-be-detected voice sample of a target speaker; preprocessing the to-be-detected voice sample; extracting acoustic features of the preprocessed to-be-detected voice sample; and inputting the acoustic features into a network reasoning module to obtain voiceprint feature vectors, wherein the network reasoning module is a network model of a single-path structure converted by a trained multi-layer network training module through re-parameterization. According to the method, the multi-branch structure is used in the training stage to achieve a better convergence effect, the multi-branch structure is re-parameterized into the single-path structure in the reasoning stage to obtain a better effect than the multi-branch structure with equivalent parameter quantity, the speed is higher, and the memory consumption is lower.

Description

technical field [0001] The present application relates to the technical field of voiceprint feature extraction, in particular to an audio feature extraction method and device based on reparameterized decoupling. Background technique [0002] Existing high-performance network structures include multi-branch structures and network components with excellent performance. Among them, the performance of the multi-branch structure can be greatly improved compared with the previous single-channel structure. Like GoogleNet, Inception, etc., all belong to the multi-channel structure. And network components with excellent performance, including depthwise separable convolution, group convolution, etc., can significantly increase network performance. However, although the multi-branch structure and high-performance components can significantly improve the performance of the model, it will eventually cause the model to slow down and consume memory during inference, which is very unfavor...

AI Technical Summary

Problems solved by technology

The core of these attempts is to train a deeper network, but there are no good results. The performance is generally not as good as the multi-branch structure, and the obtained models are often neither simple nor practical.

Images