连续语音识别中基于Dropout修正线性深度置信网络的声学模型

大词汇量连续语音识别系统中,为了增强现有声学模型的表征能力、防止模型过拟合,提出一种基于遗失策略(Dropout)修正线性深度置信网络的声学模型构建方法。该方法使用修正线性函数代替传统Logistic函数进行深度置信网络训练,修正线性函数更接近生物神经网络的工作方式,增强了模型的表征能力;同时引入Dropout策略对修正线性深度置信网络进行调整,避免节点之间的协同作用,防止网络出现过拟合。文章利用公开语音数据集进行了实验,实验结果证明了所提出的声学模型构建方法相对于传统方法的优越性。

Acoustic model based on Dropout rectified deep belief network in large vocabulary continuous speech recognition system

To improve representation ability of acoustic model and prevent over fitting in large vocabulary continuous speech recognition system, this article proposes a method of establishing the acoustic model based on Dropout rectified Deep Belief Network (DBN). This method uses rectified linear function instead of traditional Logistic function as the activation function for DBN training, and the rectified linear function that is closer to the working mode of biological neural network can improve acoustic representation ability of the model, simultaneously Dropout strategy is in-troduced to avoid the synergy between nodes and to prevent over fitting. The actual test certificate on public speech databases proves the superiority of the proposed method over the conventional one.