Multi-channel anti-crosstalk dynamic planning strategy based on feedforward memory network

Abstract

The invention discloses a multi-channel anti-crosstalk dynamic planning strategy based on a feedforward memory network. The multi-channel anti-crosstalk dynamic planning strategy specifically comprises the following steps of S1, collecting sound source information; s2, self-adaption of sound source gain is carried out; s3, a crosstalk filtering process; s4, outputting the processed sound, and obtaining a multi-channel recognition result; the self-adaptive sound source gain process provided by the invention is not limited by a specific field environment structure, hardware sound collection equipment and behaviors of a spokesman, and sound source gain which contributes to improving the recognition effect can be dynamically carried out on the acoustic signal of the spokesman in real time, so that the tuning cost of a traditional method is avoided, and the real recording process is guaranteed; through effective acoustic feature extraction and a crosstalk identification strategy model, crosstalk channels are identified and filtered in real time, and occurrence of crosstalk phenomena is significantly reduced. For an extreme field environment, adjustment parameters provided by the invention are simple and easy to use, targeted adjustment can be quickly and timely made, and the real recording effect is improved.

Description

technical field [0001] The invention relates to the field of sound processing, in particular to a multi-channel anti-crosstalk dynamic programming strategy based on a feedforward memory network. Background technique [0002] In the multi-person recording system based on speech recognition, the indoor environment structure, microphone hardware, speaker position and voice all significantly affect the on-site recognition effect. The recognition rate drops and the recognition results are wrong, which affects the user experience. Also do not have a kind of method, device or equipment that effectively solves the above problems simultaneously on the market. The self-adaptive sound source gain and anti-crosstalk method provides simple and general configuration parameters, realizes adaptation to different recording sites, comprehensively considers factors such as environment, hardware, software, and speaker mode, analyzes the acoustic information of each speaker channel in real time...

AI Technical Summary

Problems solved by technology

[0008] The present invention is to overcome the following phenomena in the prior art that are prone to occur in the recording environment of the applied speech recognition system: 1) it is difficult for a speaker with a small voice to be picked up and recognized; 2) a speaker with a loud voice is easy to be strung into other acquisition devices Cause interference; 3) Multiple speakers speaking at the same time are prone to crosstalk and drop words. 4) The on-site environment has poor noise filtering effect, resulting in a large number of interference sound sources, which leads to problems such as confusion in the recognition results. A feed-forward memory network based on The multi-channel anti-crosstalk dynamic planning strategy greatly solves the problem of low sound source loss and crosstalk of each channel, effectively improving the effect of speech recognition

Images