基于卷积神经网络的语音分离方法研究与实现

发布时间：2018-03-20 13:58

本文选题：语音分离　切入点：卷积神经网络　出处：《内蒙古大学》2017年硕士论文　论文类型：学位论文

【摘要】：随着智能手机和智能设备逐渐普及,人机语音交互技术日益得到了广泛的关注和重视。如何使人机语音交互变得像人与人之间的交流一样方便和高效成为了最近几年研究的热点问题。而人机语音交互技术中核心问题之一是语音分离,前端语音分离模块的性能对人机交互中的自动语音识别、语音自动翻译和说话者识别都有十分重要的影响。近几年随着深度学习的兴起,基于深度模型的语音分离逐渐成为了一种新的研究趋势,并且得到了较好的语音分离效果。然而,基于卷积神经网络(Convolutional Neural Network,CNN)的语音分离研究成果还非常少。并且,相比于深度神经网络(Deep Neural Network,DNN),CNN更加擅长于挖掘输入语音信号中的时空特性,能够更好地刻画语音特征。在本文中采用卷积神经网络作为建模工具进行语音分离。本文完成了单通道语音分离和多通道语音分离两个实验,其中每个实验都在相同数据集上使用卷积神经网络和深度神经网络两个模型来分别实现的。单通道语音分离是在标准语音库TIMIT上进行的实验,多通道语音分离是在CHIME3官方提供的比赛数据上进行的实验,经实验结果表明,相比于深度神经网络,卷积神经网络能够在明显降低模型复杂度和减少训练权重参数的同时,估计出具有更高主观语音质量(Perceptual Evaluation of Speech Quality,PESQ)和短时客观语音可懂度(Short Time Objective Intelligibility,STOI)的语音。
[Abstract]:With the popularity of smartphones and smart devices, The technology of human-computer speech interaction has been paid more and more attention. How to make human-computer speech interaction as convenient and efficient as the communication between people has become a hot issue in recent years. One of the core issues in interoperability is speech separation. The performance of the front-end speech separation module plays an important role in automatic speech recognition, speech translation and speaker recognition in human-computer interaction. Speech separation based on depth model has gradually become a new research trend, and has obtained better speech separation effect. However, the research results of speech separation based on convolution neural network Convolutional Neural Network (CNN) are still very few. Compared with Deep Neural Network CNN, it is better at mining temporal and spatial characteristics of input speech signals. In this paper, convolution neural network is used as a modeling tool for speech separation. In this paper, two experiments of single channel speech separation and multi channel speech separation are completed. Each experiment is implemented on the same data set using convolution neural network and depth neural network respectively. Single channel speech separation is carried out on the standard speech corpus TIMIT. Multi-channel speech separation is an experiment based on the competition data provided by CHIME3. The experimental results show that compared with the depth neural network, the convolutional neural network can significantly reduce the complexity of the model and reduce the training weight parameters at the same time. The speech with higher subjective Evaluation of Speech quality and short Time Objective intelligibilitySTOI) is estimated.
【学位授予单位】：内蒙古大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TN912.3

【相似文献】