Integration of MFCCs and CNN for Multi-Class Stress Speech Classification on Unscripted Dataset

Abstract

Stress is an interaction between individuals and their environment, where perceived threats can lead to serious consequences if prolonged and consistently linked to adverse physical and mental health outcomes. Our study explores methods for stress classification via speech, utilizing an unscripted dataset from an experimental study that was able to show the spontaneous reactions of stressed individuals. Mel-Frequency Cepstral Coefficients (MFCCs) emerge as promising speech features, adept at representing the power spectrum crucial to human auditory perception, especially in stress speech recognition. Leveraging deep learning technology, specifically Convolutional Neural Network (CNN), our research optimally combines speech features and CNN algorithms for stress classification. Despite the scarcity of publications on unscripted datasets and multi-class stress classifications, our study advocates their adoption, aiming to enhance performance metrics and contribute to research expansion. The proposed system shows that MFCCs achieve an accuracy of 95.67% in distinguishing among three stress classes (low-stress, medium-stress, and high-stress), surpassing the prior unscripted dataset study by 81.86%. This highlights the efficacy of the proposed MFCCs-CNN system in stress classification. ABSTRAK: Tekanan merupakan interaksi antara individu dan persekitaran, di mana ancaman akan membawa kepada akibat serius jika berlarutan, dan secara konsisten dikaitkan dengan kesan kesihatan fizikal dan mental yang buruk. Kajian ini mengkaji kaedah pengelasan tekanan melalui pertuturan, menggunakan set data tanpa skrip yang diperoleh daripada kajian eksperimen, iaitu mampu menunjukkan tindak balas spontan individu tertekan. Pekali Septral Frekuensi-Mel (MFCCs) muncul sebagai ciri pertuturan berpotensi, iaitu mahir dalam menunjukkan secara ringkas spektrum kuasa penting bagi persepsi pendengaran manusia, terutama ketika pengecaman pertuturan bertekanan. Memanfaatkan teknologi pembelajaran mendalam, khususnya Rangkaian Neural Lingkaran (CNN), kajian ini menggabungkan ciri pertuturan dan algoritma CNN secara optimum bagi pengelasan tekanan. Walau terdapat kekurangan penerbitan pada set data tanpa skrip dan klasifikasi tekanan pelbagai kelas, kajian ini meningkatkan penggunaannya, bertujuan bagi meningkatkan metrik prestasi dan menyumbang kepada keluasan penyelidikan. Sistem yang dicadangkan ini menunjukkan bahawa MFCC mencapai ketepatan 95.67% dalam membezakan antara tiga kelas tekanan (tekanan rendah, tekanan sederhana dan tekanan tinggi), mengatasi kajian dataset tanpa skrip terdahulu sebanyak 81.86%. Ini menunjukkan keberkesanan sistem MFCCs-CNN dalam pengelasan tekanan.