Speech Emotion Recognition in Pakistani-Accented English Language

Speech Emotion Recognition(SER) is a field of study that focuses on the automatic detection and interpretation of emotions conveyed through speech. Recognizing and understanding emotions from speech is essential in enabling machines to interact with humans in a more natural and empathetic manner. This manuscript investigates speech emotion recognition, specifically focusing on the nuances present in Pakistani-accented English - the first such initiative to the best of our knowledge. To this end, we collected a total of 2,620 audio recordings where each recording was annotated as belonging to one of the four emotion labels, i.e., anger, happiness, neutral, and sadness. For these, we trained machine learning models using Decision trees, logistic regression, support vector machines, and random forest classifiers for several standard acoustic feature sets from the openSmile toolkit such as Prosody, ISIO, ComParE, and eGeMAPS, through a 4-fold cross-validation setup. The aim here was to produce a classification benchmark for our new dataset. We report that in subject-non-independent cross-validation setup (StratifiedKFold), ComParE features with Logistic Regression classifier achieved the best result with a mean accuracy of 76.01% whereas for subject-independent setup (StratifiedGroupKFold), the same combination of acoustic features and classifier achieved an accuracy of 69.61%. We intend to release this dataset for academic research into SER and encourage fellow researchers to improve the classification baseline.