面向语音情感识别的改进长短时记忆网络研究

Speech Emotion Recognition Using LSTM with Attention-gate

为挖掘语音情感中的时序信息,提出一种基于注意力门的长短时记忆网络算法.首先提取动态语音特征保留语音中的时序信息;然后针对具有处理序列能力的长短时记忆网络,提出一种注意力门取代传统模型中的遗忘门与输入门;最后依据注意力机制来计算输出门中的历史细胞状态与候选细胞状态的权重系数,可以在减少计算复杂度的同时,提高情感识别的准确率.在eNTERFACE和GEMEP语料库上的试验表明,改进后的长短时记忆网络获得的情感识别率比传统模型分别提高了约10%和4%.此外,在相同参数下,改进后的模型所需要的训练时间更短,说明降低了计算的复杂度.

In order to mine the temporal information in speech emotion recognition,an attention-gate based long short-term memory network is proposed.Firstly,dynamic features are extracted to protect the temporal information.In terms of long short-term memory(LSTM)network with the capability to process the sequential data,a new attention-gate is proposed to replace the forgetting gate and input gate of the traditional model.Based on the attention mechanism,the weights are calculated directly for historical cell state and candidate cell state of the output gate,which not only reduce the computation complexity,but also improve the accuracy of emotion recognition.Compared with the traditional model,the experiments on the eNTERFACE and GEMEP corpus demonstrate that the accuracy of proposed method is improved by 10%and 4%,respectively.Furthermore,the improved model takes less training time with the same parameters,which means the computation complexity is reduced.