Auditory selection method and device based on memory and attention model

Abstract

The invention belongs to the technical field of speech separation, and in particular relates to an auditory selection method and device based on memory and attention models. It aims to solve the problems of the arrangement of supervised labels, the uncertain number of speaker aliasing and the fixed dimension of memory units in the prior art. The invention provides an auditory selection method based on a memory and attention model, which includes encoding the original speech signal into a time-frequency matrix, encoding and transforming the time-frequency matrix, converting it into a speech vector, and using a long-term memory unit to store speech People and their corresponding speech vectors, obtain the speech vectors of the target speaker, and separate the target speech from the original speech signal through the attention selection model. The method provided by the invention can separate the target speech from the original speech signal without fixing or specifying the number of speakers.

Description

technical field [0001] The invention belongs to the technical field of speech separation, and in particular relates to an auditory selection method and device based on memory and attention models. Background technique [0002] In recent years, with the rapid development of electronic equipment and artificial intelligence, human-computer voice interaction, as an important part of the field of artificial intelligence, has become increasingly important, and human-computer voice interaction has been widely used in real life. Human-computer voice interaction is the machine recognition and analysis to extract the semantic feature information of the voice signal, compare it with the semantic features in the standard information base, and output the corresponding text or convert it into the output result we want. However, in practical applications, there are a lot of interference in the real environment, and the process of machine recognition, analysis and extraction of semantic fea...

AI Technical Summary

Problems solved by technology

[0005] In order to solve the above problems in the prior art, that is, in order to solve the problems of the arrangement of supervised labels, the uncertain number of speaker aliasing and the fixed dimension of the memory unit in the prior art, an aspect of the present invention provides a method based on memory and attention. Auditory selection methods for force models, including:

Images