Using Speaker-Specific Emotion Representations in Wav2vec 2.0-Based Modules for Speech Emotion Recognition

Abstract

Speech emotion recognition is essential for frictionless human-machine interaction, where machines respond to human instructions with context-aware actions. The properties of individuals’ voices vary with culture, language, gender, and personality. These variations in speaker-specific properties may hamper the performance of standard representations in downstream tasks such as speech emotion recognition (SER). This study demonstrates the significance of speaker-specific speech characteristics and how considering them can be leveraged to improve the performance of SER models. In the proposed approach, two wav2vec-based modules (a speaker-identification network and an emotion classification network) are trained with the Arcface loss. The speaker-identification network has a single attention block to encode an input audio waveform into a speaker-specific representation. The emotion classification network uses a wav2vec 2.0-backbone as well as four attention blocks to encode the same input audio waveform into an emotion representation. These two representations are then fused into a single vector representation containing emotion and speaker-specific information. Experimental results showed that the use of speaker-specific characteristics improves SER performance. Additionally, combining these with an angular marginal loss such as the Arcface loss improves intra-class compactness while increasing inter-class separability, as demonstrated by the plots of t-distributed stochastic neighbor embeddings (t-SNE). The proposed approach outperforms previous methods using similar training strategies, with a weighted accuracy (WA) of 72.14% and unweighted accuracy (UA) of 72.97% on the Interactive Emotional Dynamic Motion Capture (IEMOCAP) dataset. This demonstrates its effectiveness and potential to enhance human-machine interaction through more accurate emotion recognition in speech.