Impact of Lightweight Machine Learning Models for Speech Emotion Recognition

Abstract

Speech Emotion Recognition (SER) is the operation or series of steps used for the identification of human emotions from verbal expressions. SER is a challenging task, and extensive reliance has been placed on models which use audio features in developing well-performing classifiers. Practitioners rely more on the power of the deep learning models, but even lighter and more interpretable ML models can achieve performance that was achieved by DL-based models with close values of accuracy, f-score, precision, and recall. Lighter machine learning-based models trained over a few handcrafted features, such as pitch, harmonics, speech energy, and pause, are able to achieve a performance comparable to the current deep learning-based state-of-the-art method for emotion recognition. To increase the accuracy of the models, we also used the features such as MFCC in the feature extraction process. In this paper, we implement the lightweight interpretable machine learning models namely, Random Forest, Gradient Boosting, Support Vector Machines, Naive Bayes, and Logistic Regression, for speech emotion detection. The ensembling of these models is done to find out the best combination for better accuracy. We developed an application for improving the customer-care services by classifying and prioritizing the feedback based on the emotions of the voice data.KeywordsSpeech emotion recognitionRandom forestGradient boostingSupport vector machinesNaive BayesAnd logistic regression