A review on speech emotion recognition: a survey, recent advances, challenges
Аннотация
Speech emotion recognition (SER) has emerged as a crucial area of research, driven by the growing need for more natural and effective human-computer interaction. Emotions play a vital role in human communication, and equipping machines with the ability to recognize and respond to human emotions has numerous practical applications in fields such as education, healthcare, customer service, virtual assistants, and more. Recent breakthroughs in deep learning (DL) technologies have significantly enhanced the capabilities of systems designed for speech-related tasks. Notably, SER has undergone transformative progress due to the development of sophisticated neural network architectures. This paper presents a comprehensive survey of recent work in the domain of SER, focusing on studies that utilize ML and DL techniques. It provides a detailed comparison of state-of-the-art models, including convolutional neural networks (CNNs), long short-term memory (LSTM) networks, Transformer-based models, and hybrid architectures. Additionally, this study discusses the critical components that contribute to SER performance, such as the selection and quality of speech emotion datasets, feature extraction techniques (e.g., Mel-frequency cepstral coefficients (MFCCs), spectrograms, prosodic features), and model evaluation metrics. Furthermore, this review highlights the existing limitations and challenges in current SER research, including issues related to dataset diversity, emotion representation, speaker and language dependency, and real-time implementation. By synthesizing the findings from various recent studies, this paper aims to provide valuable insights for researchers to identify potential directions for future work to address the unresolved issues in this evolving field.