基于DBM-LSTM的多特征语音情感识别

为增强不同情感特征的融合程度和语音情感识别模型的鲁棒性,提出一种神经网络结构DBM-LSTM用于语音情感识别。利用深度受限玻尔兹曼机的特征重构原理将不同的情感特征进行融合;利用长短时记忆单元对短时特征进行长时建模,增强语音情感识别模型的鲁棒性;在柏林情感语音数据库上进行分类实验。研究结果表明,与传统识别模型相比,DBM-LSTM网络结构更适用于多特征语音情感识别任务,最优识别结果提升11%。

Multi-feature speech emotion recognition based on DBM-LSTM

To enhance the fusion degree of different emotional features and the robustness of speech emotion recognition model,a neural network structure of deep-restricted Boltzmann machine and long short term memory(DBM-LSTM)for speech emotion recognition was proposed.The principle of feature reconstruction based on DBM was used to fuse the different emotional features.LSTM was used to model short-term features for a long-term to enhance the robustness of speech emotion recognition model.The classification experiments were carried out on EMO-DB database.The results show that,compared with traditional recognition model,DBM-LSTM network structure is more suitable for multi-feature speech emotion recognition tasks.The optimal recognition results can be improved by 11%.