基于BLSTM的语音脑电多阶段结合目标语音提取  

Target speech separation with BLSTM-based multi-stage combination of speech and EEG

作　　者：裴意静 贾海蓉[1] 吉陈果 段淑斐[1] PEI Yi-jing;JIA Hai-rong;JI Chen-guo;DUAN Shu-fei(College of Electronic Information and Optical Engineering,Taiyuan University of Technology,Jinzhong 030600,China)

机构地区：[1]太原理工大学电子信息与光学工程学院,山西晋中030600

出　　处：《计算机工程与设计》2025年第2期479-484,共6页Computer Engineering and Design

基　　金：国家自然科学基金项目(12004275);山西省自然科学基金项目(20210302123186)。

摘　　要：针对单通道混合音频目标语音提取困难的问题,依据脑电信号携带能直接反映注意和感知的信息,提出一种将语音与脑电信号结合进行目标语音提取的基于BLSTM的AESE模型。采集语音诱发的EEG信号构建数据库,进行脑电多特征的目标语音提取可行性研究。选取脑电特征TS16R和语音频谱图特征通过AESE进行多阶段结合,生成含丰富声学上下文信息的脑电特征掩码,提升重构纯净语音的效果。实验结果表明,AESE从多说话人混合音频及带噪语音中有效提取目标语音,为基于脑电的目标语音提取提供了一种思路和方法。Aiming at the difficulty of target speech extraction in single-channel mixed audio,a BLSTM-based AESE model was proposed to extract target speech based on EEG signals,directly reflecting attention and perception information.The speech induced EEG signals were collected to build a database,and the feasibility of multi-feature target speech extraction was studied.EEG feature TS16R and speech spectrogram feature were selected for multi-stage combination through AESE to generate EEG feature mask with rich acoustic context information,and improve the effect of reconstructing pure speech.Experimental results show that AESE can effectively extract target speech from multi-speaker mixed audio and noisy speech,which provides an idea for target speech extraction based on EEG.

关 键 词：目标语音提取 脑电信号 数据采集 特征提取 掩码 特征融合 多阶段结合 

分 类 号：TN912.3[电子电信—通信与信息系统]