Hi, I am Nguyen Minh Nhut.

Multimodal learning has been shown to enhance classification outcomes in speech emotion recognition (SER). Despite this advantage, multimodal approaches in SER often face key challenges, including limited robustness to uncertainty, difficulty generalizing across diverse emotional contexts, and inefficiencies in integrating heterogeneous modalities. To overcome these constraints, we propose a multimodal emotion recognition architecture, named FleSER, which leverages dynamic fuzzy membership and attention-based fusion. Unlike most previous SER studies that apply fuzzy logic at the decision level, FleSER introduces a feature-level, rule-based dynamic fuzzy membership mechanism that adaptively refines modality representations prior to fusion. The FleSER architecture leverages audio and textual modalities, employing self-modality and cross-modality attention mechanisms with the α interpolation to capture complementary emotional cues. The α interpolation-based feature fusion mechanism adaptively emphasizes the more informative modality across varying contexts, ensuring robust multimodal integration. This comprehensive design enhances recognition accuracy. We evaluate FleSER on multiple benchmark datasets, surpassing previous state-of-the-art (SOTA) approaches and demonstrating superior effectiveness in emotion recognition. Ablation studies further validate the effectiveness of each key component, including unimodal and multimodal input effectiveness, fuzzy membership functions, fusion strategies, and the projection dimension, on the performance of the FleSER architecture.

Speech emotion recognition (SER) plays a crucial role in human-computer interaction, enhancing numerous applications such as virtual assistants, healthcare monitoring, and customer support by identifying and interpreting emotions conveyed through spoken language. While single-modality SER systems demonstrate notable simplicity and computational efficiency, excelling in extracting critical features like vocal prosody and linguistic content, there is a pressing need to improve their performance in challenging conditions, such as noisy environments and the handling of ambiguous expressions or incomplete information. These challenges underscore the necessity of transitioning to multi-modal approaches, which integrate complementary data sources to achieve more robust and accurate emotion detection. With advancements in artificial intelligence, especially in neural networks and deep learning, many studies have employed advanced deep learning and feature fusion techniques to enhance SER performance. This review synthesizes comprehensive publications from 2020 to 2024, exploring prominent multi-modal fusion strategies, including early fusion, late fusion, deep fusion, and hybrid fusion methods, while also examining data representation, data translation, attention mechanisms, and graph-based fusion technologies. We assess the effectiveness of various fusion techniques across standard SER datasets, highlighting their performance in diverse tasks and addressing challenges related to data alignment, noise management, and computational demands. Additionally, we explore potential future directions for enhancing multi-modal SER systems, emphasizing scalability and adaptability in real-world applications. This survey aims to contribute to the advancement of multi-modal SER and to inform researchers about effective fusion strategies for developing more responsive and emotion-aware systems.

Multimodal Speech Emotion Recognition (SER) offers significant advantages over unimodal approaches by integrating diverse information streams such as audio and text. However, effectively fusing these heterogeneous modalities remains a significant challenge. We propose CemoBAM, a novel dualstream architecture that effectively integrates the Heterogeneous Graph Attention Network (CH-GAT) with the Cross-modal Convolutional Block Attention Mechanism (xCBAM). In CemoBAM architecture, the CH-GAT constructs a heterogeneous graph that models intra- and inter-modal relationships, employing multi-head attention to capture fine-grained dependencies across audio and text feature embeddings. The xCBAM enhances feature refinement through a cross-modal transformer with a modified 1D-CBAM, employing bidirectional cross-attention and channel-spatial attention to emphasize emotionally salient features. The CemoBAM architecture surpasses previous state-of-the-art (SOTA) methods by 0.32% on IEMOCAP and 3.25% on ESD datasets. Comprehensive ablation studies validate the impact of Top-K graph construction parameters, fusion strategies, and the complementary contributions of both modules. The results highlight CemoBAM’s robustness and potential for advancing multimodal SER applications.