基于注意力机制和RRM机制的实时语音英译去噪系统研究  

Research on real-time speech English translation denoising system based on attention mechanism and RRM mechanism

作　　者：唐竹 TANG Zhu(Xi’an Haitang Vocational College,Xi’an 710038,China)

机构地区：[1]西安海棠职业学院,西安710038

出　　处：《自动化与仪器仪表》2024年第12期267-271,275,共6页Automation & Instrumentation

摘　　要：为提升目前实时语音英译去噪系统的去噪能力和实时性,研究设计了一种结合注意力机制和重复完善机制(Repetitive Refinement Mechanisms, RRM)机制的网络模型。该网络通过分离混合语音信号中目标说话人的语音实现非目标说话人人声数据去除,同时通过SF模块的引入克服Spex+网络在长序列语音信息中无法获取全局信息的缺点。数据规模从100个增加到1 000个,该网络模型的感知语音质量测度和短时语音可懂度数值均表现先固定不变,随后逐渐升高的变化趋势。该网络模型利用RRM机制获得目标说话人提取器和说话人编码器的初始输出得到注册语音特征向量,同时在Transformer结构图中采用多头注意力机制,极大提升语音数据的理解和处理水平。该实时语音英译去噪系统在满足实时性和去噪能力的基础上,提升线上会议的语音质量和可懂性。To improve the denoising ability and real-time performance of current real-time speech translation denoising systems,a network model combining Repetitive Refinement Mechanisms and repetitive refinement mechanisms(RRM)was designed.The network can remove the non-target speaker's voice data by separating the target speaker's voice from the mixed speech signal.Meanwhile,the SF module is introduced to overcome the shortcoming that Spex+network cannot obtain global information in long sequence speech information.Data size increased from 100 to 1,000,The network model's PESQ and STOI,The values of STOI are all fixed at first and then gradually increasing.In this network model,RRM mechanism is used to obtain the initial output of target speaker extractor and speaker encoder to obtain the registered speech feature vector.Meanwhile,multi-head attention mechanism is used in Transformer structure diagram to greatly improve the understanding and processing level of speech data.On the basis of satisfying the real-time and denoising ability,this system can improve the speech quality and intelligibility of online conference.

关 键 词：注意力机制 RRM 实时语音 英译去噪 时域 

分 类 号：TN912.3[电子电信—通信与信息系统] TP29[电子电信—信息与通信工程]