An integrated deep learning model combining 1D-UNet and LSTM for speech emotion classification

Speech emotion detection is crucial for various applications, including audio monitoring, clinical research, e-learning, entertainment, lie detection, gaming, and customer service centres. Classifying sentimental states from speech data necessitates advanced methods to achieve optimal predictive performance. Deep learning (DL) approaches have revealed efficacy in recording high-level traits of speech data and addressing categorisation challenges. This investigation aims to predict different sentimental states using speech signal exploration and DL frameworks. Specifically, we introduce two network types: (i) the 1D-UNet framework, which consists solely of deep one-dimensional CNN (1D-CNN), and (ii) the 1D-UNet-long short-term memory (1D-UNet-LSTM), which integrates the 1D-UNet construction with the LSTM recurrent network. Three databases namely RML, RAVDESS, IEMOCAP and SAVEE are utilised in the empirical analyses to assess the efficacy of the recommended networks. The 1D-UNet-LSTM outperformed the 1D-UNet framework for speech emotion detection in all databases. The 1D-UNet-LSTM network obtained 95.90%, 97.85%, 90.11%, and 86.23% mean accuracy rates achieved for RML, RAVDESS, SAVEE, and IEMOCAP databases, accordingly; whereas, the 1D-UNet framework yielded 95.19%, 96.77%, 88.82%, and 83.51% mean precision rates were achieved for RML, RAVDESS, SAVEE, and IEMOCAP databases, accordingly. The recommended networks demonstrate notable efficiency, resulting in improved accuracy rates compared to previously published outcomes when tested on the relevant databases.