多语音和深度学习的对话机器人语音增强技术研究  

Research on Speech Enhancement Technology for Dialogue Robot Based on Multi Speech and Deep Learning

作　　者：王小莉[1] WANG Xiaoli(Xian Yang Vocational&Technical College,Xianyang,Shannxi 712000,China)

机构地区：[1]咸阳职业技术学院,陕西咸阳712000

出　　处：《自动化与仪器仪表》2023年第12期173-177,181,共6页Automation & Instrumentation

基　　金：《课程思政视域下陕西地域特色文化融入高职英语教学的路径研究》(SGH22Y1667);《基于国际化人才培养的外语课程体系构建研究》(2023HZ0974);《“一带一路”背景下高校外语专业国别与区域研究及人才培养模式的探索》(2023HZ1014)。

摘　　要：为了提升多语音对话机器人在英语语音交流过程中的去噪性能,研究在深度神经网络(Deep Neural Network,DNN)模型的语音增强算法基础上,提出一种时频掩码优化的二阶段英语语音增强算法。该算法通过两阶段网络提升多语音对话机器人英语语音增强性能,设置目标函数为时频掩码优化函数,利用相位信息系数和增益系数提升网络去噪性能。特征拼接系统的两种评价指标均高于单一特征系统。相较于其他类型的深度学习算法,时频掩码优化的二阶段英语语音增强算法具有更理想的客观英语语音质量评价性能。当信噪比(Signal-Noise Ratio,SNR)为5 dB时,时频掩码优化的二阶段英语语音增强算法的STOI和PESQ分别为0.752和2.865,数值均为最高。所构建的二阶段英语语音增强算法能适应多种类型的噪音环境,具有较强的英语语音噪声抑制作用,能保留较为完整的英语语音结构信息,丰富了英语语音增强理论,可应用于智能设备的英语语音增强领域。To improve the noise reduction performance of multi-speech dialogue robots in the process of English speech communication,a two-stage English speech enhancement algorithm based on the Deep Neural Network(DNN) model is proposed.The algorithm improves the English speech enhancement performance of multi-voice dialogue robot through two-stage network,sets the objective function of time-frequency mask optimization function,and uses the phase information coefficient and gain coefficient to improve the network denoising performance.The two evaluation indexes of feature splicing system are higher than those of single feature system.Compared with other types of deep learning algorithms,the two-stage English speech enhancement algorithm with time-frequency mask optimization has better performance of objective English speech quality evaluation.When the Signal-Noise Ratio(SNR) is 5 dB,the STOI and PESQ of the two-stage English speech enhancement algorithm with time-frequency mask optimization are 0.752 and 2.865,respectively,with the highest values.The proposed two-stage English speech enhancement algorithm can adapt to various types of noise environment,has a strong English speech noise suppression effect,can retain a relatively complete English speech structure information,enrich the English speech enhancement theory,and can be applied to the English speech enhancement field of intelligent devices.

关 键 词：CNN 多语音对话机器人 英语语音增强 SNR 

分 类 号：TP29[自动化与计算机技术—检测技术与自动化装置] TN912.35[自动化与计算机技术—控制科学与工程]