Deep feature representations and fusion strategies for speech emotion recognition from acoustic and linguistic modalities: A systematic review

Emotion Recognition (ER) has gained significant attention due to its importance in advanced human-machine interaction and its widespread real-world applications. In recent years, research on ER systems has focused on multiple key aspects, including the development of high-quality emotional databases, the selection of robust feature representations, and the implementation of advanced classifiers leveraging AI-based techniques. Despite this progress in research, ER still faces significant challenges and gaps that must be addressed to develop accurate and reliable systems. To systematically assess these critical aspects, particularly those centered on AI-based techniques, we employed the PRISMA methodology. Thus, we include journal and conference papers that provide essential insights into key parameters required for dataset development, involving emotion modeling (categorical or dimensional), the type of speech data (natural, acted, or elicited), the most common modalities integrated with acoustic and linguistic data from speech and the technologies used. Similarly, following this methodology, we identified the key representative features that serve as critical emotional information sources in both modalities. For acoustic, this included those extracted from the time and frequency domains, while for linguistic, earlier embeddings and the most common transformer models were considered. In addition, Deep Learning (DL) and attention-based methods were analyzed for both. Given the importance of effectively combining these diverse features for improving ER, we then explore fusion techniques based on the level of abstraction. Specifically, we focus on traditional approaches, including feature-, decision-, DL-, and attention-based fusion methods. Next, we provide a comparative analysis to assess the performance of the approaches included in our study. Our findings indicate that for the most commonly used datasets in the literature: IEMOCAP and MELD, the integration of acoustic and linguistic features reached a weighted accuracy (WA) of 85.71% and 63.80%, respectively. Finally, we discuss the main challenges and propose future guidelines that could enhance the performance of ER systems using acoustic and linguistic features from speech.