Detecting hate speech on social media is challenging, particularly in low-resourced languages like Malayalam, due to the scarcity of annotated data. To address this challenge, we introduce a new multiclass dataset for hate speech in the Malayalam language, sourced from YouTube. The study benchmarks the performance of machine learning classifiers for the classification of hate and non-hate speech, in both binary and multiclass classification tasks, using audio features alone. The Random Forest Classifier model performed exceptionally well in binary classification, achieving a macro accuracy of 0.93 and an F1-score of 0.93. Ablation studies conducted with other classifiers, such as Logistic Regression, Support Vector Machines, and Naive Bayes, registered accuracies around 0.85 and macro F1-scores of 0.85. In multiclass classification, the Random Forest model excelled with an accuracy of 0.8289, a macro accuracy of 0.72, and an F1-score of 0.74, outperforming all other models tested in the ablation study. These results demonstrate the effectiveness of the Random Forest Classifier in contributing to a safer online environment by reliably detecting hate speech in Malayalam.