Speech Emotion Recognition

Speech emotion recognition (SER) estimates emotion from vocal cues such as pitch, rhythm, and voice quality, often predicting categories or continuous dimensions like valence and arousal. Approaches range from handcrafted acoustic features and CNN or recurrent models to Transformers and self-supervised speech encoders; some combine speech with transcripts, conversational context, or audio-language models. Systems are commonly tested on datasets such as IEMOCAP, RAVDESS, and MELD, but performance can fall across speakers, languages, recording conditions, and natural versus acted speech. A central challenge is that emotion labels are subjective and models may rely on words or dataset shortcuts rather than vocal expression, making robust evaluation and real-world reliability open concerns.