Speaker model compression system and method based on double-layer knowledge distillation

Abstract

The invention discloses a speaker model compression system and method based on double-layer knowledge distillation, and belongs to the technical field of implementation modes of stereo matching algorithms. The method comprises the following steps of: guiding a student network to simulate segment-level speaker representation (speaker representation) of a teacher network by Embardment layer knowledge distillation, and capturing basic distribution of features of each speaker by the segment-level speaker representation; knowledge distillation of the Logit layer guides a student network to simulate speaker posterior probability distribution of a teacher network, and similarity among speaker categories is utilized. According to the method, a hierarchical structure of speaker representation distribution is migrated from a teacher network. According to the invention, the problems in the prior art that a student network cannot realize small difference in speakers and large difference between speakers, and the accuracy of a verification system for the same speaker and different speakers is low are solved.

Description

technical field [0001] The invention belongs to the technical field of model compression based on double-layer knowledge distillation technology, and in particular relates to a speaker model compression system and method based on double-layer knowledge distillation. Background technique [0002] In recent years, with the increasing abundance of computing resources and data resources. Machine learning based on deep neural networks has significantly improved the accuracy of speaker recognition systems. For situations where the network connection is unavailable or people are concerned about the leakage of personal privacy, people hope to use speaker recognition technology locally on embedded devices such as mobile phones. This kind of speaker recognition system running on embedded terminals has a higher level of security. However, existing speaker recognition techniques rely on deep neural networks, and their expensive computation and large memory footprint hinder their deploy...

AI Technical Summary

Problems solved by technology

[0004] The purpose of the present invention is to provide a speaker model compression system and method based on double-layer knowledge distillation, to solve the problem that the middle school student network in the prior art cannot achieve small intra-speaker differences and large inter-speaker differences, and the same speaker Problems with lower accuracy of verification systems with different speakers

Images