【摘要】 语音是信息传递的载体。情感作为语音携带的一种突出的信息,左右了语言的释义。在对现有语音情感识别方法进行研究的基础上,提出一个深度迁移网络——基于注意力机制的长短时动态对抗适配网络(Attention-based LSTM Dynamic Adversarial Adaptation Networks,LSTM-TF-at-DAAN)进行跨库语音情感识别。在语音情感识别领域广泛使用的eNTERFACE语音库和EMO-DB语音库上进行实验,并将实验结果与采用一般迁移学习方法的实验结果进行对比,发现LSTM-TF-at-DAAN提升了5.37%的识别准确率,为深度迁移学习应用于跨库语音情感识别提供了可行性证明。更多还原

【Abstract】 Voice is the carrier of information transmission. Emotion, as a kind of prominent information carried by voice, influences the interpretation of language. Based on the research of existing speech emotion recognition methods, a deep transfer network—Attention-based LSTM Dynamic Adversarial Adaptation Networks(LSTM-TF-at-DAAN) is proposed for cross corpus speech emotion recognition. The experiment is carried out on the eNTERFACE and EMO-DB library, which are widely used in the field of speech emotion recognition. The experimental results are compared with the experimental results using the general transfer learning method, and it is found that LSTM-TF-at-DAAN improves the recognition accuracy by 5.37%, which provides a feasibility proof for the application of deep transfer learning to cross-database speech emotion recognition.更多还原