Emotion Recognition in Kashmiri Speech: Evaluating Coefficient-Based Acoustic Features Using Bidirectional LSTM Networks

This study pioneers emotion recognition in the Kashmiri language by developing the first comprehensive audio corpus for this linguistically underresourced language. Three sets of acoustic features—Chroma Features, Linear Predictive Coding (LPC), and Mel-Frequency Cepstral Coefficients (MFCCs)—are carefully looked at by a number of high-performance classification models. Experimental findings demonstrate MFCCs achieve superior performance with an accuracy of 83%, significantly outperforming LPC (63%) and Chroma features (41%). This highlights the efficacy of MFCCs in capturing fine-grained acoustic patterns crucial for precise emotional differentiation. Validation using confusion matrices and Receiver Operating Characteristic (ROC) curve analysis shows that MFCCs are very good at separating features that are specific to emotions. These results underscore the critical role of optimized feature selection in developing accurate emotion recognition models for low-resource languages. This work establishes a foundational dataset and elucidates key insights into feature performance, advancing computational research and emotion recognition methodologies in Kashmiri and similarly underresourced languages.