A Lightweight ECAPA-TDNN Framework for Persian Speech Emotion Recognition

Speech emotion recognition (SER) plays an important role in affective computing and human–computer interaction. However, developing robust SER systems for low-resource languages such as Persian remains challenging due to limited annotated data, class imbalance, inter-speaker variability, and acoustic overlap among emotion categories. This study proposes a compact transfer-learning framework for Persian SER based on the ECAPA-TDNN architecture together with a novel Emotion-Aware Loss. By incorporating inter-emotion similarity into the optimization objective, the proposed method improves discrimination between acoustically similar emotion categories without modifying the underlying network architecture. Experiments on the speaker-independent ShEMO benchmark demonstrate that the proposed framework achieves ۸۳.۰۴% Accuracy, ۸۸.۵۴% Unweighted Average Recall (UAR), and ۸۸.۲۴% F۱-score, consistently outperforming ECAPA-TDNN baselines trained with cross-entropy, standard Additive Angular Margin Softmax (AAM-Softmax), and fixed similarity weighting. The results demonstrate that similarity-aware optimization improves emotional representation learning while preserving the compact architecture of ECAPA-TDNN, making the proposed framework an effective approach for Persian speech emotion recognition.

کلیدواژه ها:

ECAPA-TDNN ، Emotion-Aware Loss ، Speech emotion recognition ، ShEMO ، SpeechBrain