Dual-Waveform Emotion Recognition Model for Conversations

Emotion recognition in conversations has broad application prospects for the current era of information explosion. Existing methods fail to utilize the detailed emotional information in speech. In this paper, we introduce a dual-waveform emotion recognition model for conversations. We 1) design a waveform-attention module capturing emotion features from source waveform and synthetic waveform; 2) use efficacy co-efficient mechanism for fine-grained multi-modal information fusion; 3) present a new dialogue emotion detection module. Extensive experiments show that our method achieves the state-of-the-art performance on IEMOCAP and SEMAINE datasets.