基于DNN-LSTM的VAD算法

DNN-LSTM based VAD algorithm

基于深度神经网络（deep neural network,DNN）的语音活动性检测（voice activity detection,VAD）忽略了声学特征在时间上的相关性,在带噪环境下性能会明显下降。该文提出了一种基于深度神经网络和长短时记忆单元（long-short term memory,LSTM）的混合网络结构应用于VAD问题。进一步对语音帧的动态信息加以分析利用,同时结合DNN-LSTM结构使用一种基于上下文信息的代价函数用于网络训练。实验语料基于TIDIGITS语音库,使用Noisex-92噪声库加噪。实验结果表明：在不同噪声环境下基于DNN-LSTM的VAD方法比基于DNN的VAD方法性能更好,新的代价函数比传统的代价函数更适用于该文提出的算法。

Voice activity detection（VAD）algorithms based on deep neural networks（DNN）ignore the temporal correlation of the acoustic features between speech frames which significantly reduces the performance in noisy environments.This paper presents a hybrid deep neural network with long-short term memory（LSTM）for VAD analyses which utilizes dynamic information from the speech frames.A context information based cost function is used to train the DNN-LSTM network.The noisy speech corpus used here was based on TIDIGITS and Noisex-92.The results show that the DNN-LSTM based VAD algorithm has better recognition accuracy than DNN-based VAD algorithms in noisy environment which shows that this cost function is more suitable than the traditional cost function.