基于RefineNet的端到端语音增强方法

为提高神经网络对语音信号时域波形的直接处理能力,提出了一种基于RefineNet的端到端语音增强方法.本文构建了一个时频分析神经网络,模拟语音信号处理中的短时傅里叶变换,利用RefineNet网络学习含噪语音到纯净语音的特征映射.在模型训练阶段,用多目标联合优化的训练策略将语音增强的评价指标短时客观可懂度(Short-time objective intelligibility,STOI)与信源失真比(Source to distortion ratio,SDR)融入到训练的损失函数.在与具有代表性的传统方法和端到端的深度学习方法的对比实验中,本文提出的算法在客观评价指标上均取得了最好的增强效果,并且在未知噪声和低信噪比条件下表现出更好的抗噪性.

RefineNet-based End-to-end Speech Enhancement

In order to improve the direct processing ability of the neural network to the time domain waveform of speech signal,this paper proposes an end-to-end speech enhancement method based on RefineNet.To simulate the short-time Fourier transform,a time-frequency analysis neural network is used in speech signal processing and the RefineNet is used to learn the feature mapping of noisy speech to clean speech.The speech enhancement evaluation metric short-time objective intelligibility(STOI)and source to distortion ratio(SDR)are integrated into the training loss function in the model training phase by using the multi-objective joint optimization training strategy.Experiments show that the proposed method consistently outperforms conventional methods and end-to-end deep learning methods on objective evaluation metric and has better noise immunity under unseen noise and low SNR conditions than other methods.