Pre-attentive speech signal processing with adaptive routing for emotion recognition

Emotion recognition from speech is essential for various applications in human–computer interaction, customer service, healthcare, and entertainment. However, developing robust and reproducible Speech emotion recognition (SER) systems is challenging due to the complexity of emotions and variability in speech signals. In this paper, we first define the concept of reproducibility in the context of deep learning models. We then introduce SpeechNet, a novel deep learning model designed to enhance reproducibility and robustness in SER. SpeechNet integrates multiple advanced components: speech recall, speech attention, and speech signal refinement modules to effectively capture temporal dependencies and emotional cues in speech signal. Additionally, it incorporates a pre-attention mechanism and a modified routing technique to improve feature emphasis and processing efficiency. We also explore effective acoustic feature fusion technique. Extensive experiments on several benchmark datasets demonstrate that the SpeechNet model achieves better performance and reproducibility compared to existing models. By addressing reproducibility and robustness, SpeechNet sets a new standard in SER, facilitating reliable and practical applications.