Transformer-Based Multi-Microphone Fusion Noise Reduction Technique aims to reduce speech noise from various microphones(acoustic,optical,bone-conduction)and improve the signal-to-noise ratio to adapt to different environments.To address the issue of suboptimal feature extraction from different channels in traditional multi-microphone fusion noise reduction algorithms,a Transformer-based approach is proposed.This algorithm consists of three main steps:Firstly,employing multi-head attention mechanism to enable each channel to learn different weights for better spatial feature learning between channels;Secondly,feeding the obtained channel features and original features into a Transformer model to generate time-domain filters;Finally,obtaining enhanced speech data for each channel through one-dimensional convolution operation.Experimental results demonstrate that the proposed algorithm achieves superior noise reduction performance.
万方数据
维普
