基于改进FCN和PSC的语言学习对话系统语音去噪及增强方法  

A Speech Denoising and Enhancement Method for Language Learning Dialogue System Based on Improved FCN and PSC

作　　者：张洁[1] ZHANG Jie(Xi'an Fanyi University,Xi’an 710105,China)

机构地区：[1]西安翻译学院,西安710105

出　　处：《自动化与仪器仪表》2024年第11期163-166,171,共5页Automation & Instrumentation

基　　金：陕西省教育科学“十四五”规划2021年度课题《语言服务背景下应用型本科院校英语口译实训体系的构建研究》(SGH21Y0451)。

摘　　要：随着语音识别技术的发展,对话系统在语言学习领域也得到了广泛应用。为避免语音信号受各种噪声干扰,导致对话系统的性能下降。研究提出了一种基于改进全卷积神经网络和相位谱补偿的语言学习对话系统语音去噪及增强方法。研究利用信噪比对相位谱补偿中的补偿因子进行改进,然后将稠密卷积网络结构引入到全卷积神经网络中对其进行改进。实验结果表明,研究设计的语音增强方法在不同输入信噪比情况下相较于原始带噪语音,分段信噪比值提升了30.54%、23.55%、18.45%。且在不同场景下,其感知语音质量评估分数均在2.5以上。这说明研究设计方法能够实现较快地计算效率下的语音增强及去噪,帮助学生学习英语口语。With the development of speech recognition technology,dialogue systems have also been widely applied in the field of language learning.To avoid the performance degradation of the dialogue system caused by various noise interferences on speech signals.A study proposes a speech denoising and enhancement method for language learning dialogue systems based on improved fully convolutional neural networks and phase spectrum compensation.Research on improving the compensation factor in phase spectrum compensation using signal-to-noise ratio,and then introducing a dense convolutional network structure into a fully convolutional neural network to improve it.The experimental results show that the designed speech enhancement method has improved the segmented signal-to-noise ratio by 30.54%,23.55%,18.45%,and 9.45% compared to the original noisy speech under different input signal-to-noise ratios.And in different scenarios,their perceived speech quality evaluation scores are all above 2.5.This indicates that the research design method can achieve speech enhancement and denoising with fast computational efficiency,helping students learn English speaking.

关 键 词：语言学习 对话系统 语音增强 语音去噪 稠密卷积网络 

分 类 号：TN912[电子电信—通信与信息系统] TP39[电子电信—信息与通信工程]