Frame-Level Audio Hate Speech Detection for Kannada Language

Detecting hate speech in audio has become increasingly challenging due to the increasing use of internet platforms and digital communication. Through this study, we develop an audio-based speech classifier to facilitate the detection of hate speech in the Kannada language. We present an approach to classifying hate speech at the frame level by extracting audio features such as Mel-Frequency Cepstral Coefficients (MFCCs), spectral bandwidth, spectral contrast, and chroma features. Furthermore, we present a custom Kannada hate speech dataset to address the scarcity of resources for hate speech studies in the Kannada language. We collected over 40 minutes of audio samples from YouTube and X (formerly Twitter). Our experiments show that an optimized XGBoost model achieved an accuracy of 73% on the custom dataset for frame-level classification. We also propose a cascading classifiers approach with two classifiers to exploit the locality of hate speech that improves the accuracy to 77%. Finally, we benchmark the proposed model against Logistic Regression, SVM, and XGBoost models.