Deep dive into speech privacy: convolutional neural networks in sensitive content detection

Abstract

Sensitive speech, encompassing hate speech, offensive language, and discriminatory remarks, has become a pressing issue in today's digital landscape. This research focuses on exploring the efficacy of Convolutional Neural Networks in sensitive speech detection and classification. CNNs have shown exceptional performance in various natural language processing tasks, capable of capturing contextual information through learned features. The study encompasses a comprehensive analysis of audio data preprocessing techniques, designing and training CNN, Bi-LSTM, and LSTM architectures, and evaluating their performance using appropriate metrics. Various data augmentation techniques are employed to enhance the diversity and robustness of the dataset. The CNN, Bi-LSTM, and LSTM architecture are carefully designed, considering the specific requirements of sensitive speech detection. Convolutional layers are employed to extract important features while pooling layers down to sample the extracted representations to capture essential information efficiently. The trained Bi-LSTM and LSTM models are evaluated using appropriate metrics, such as accuracy, loss,val_loss, and val_accuracy. Here the highest accuracy 98.68% achieved by the CNN Base model and the lowest loss 10.43% is also achieved by the CNN base model. However, the Bi-LSTM and LSTM model shows the comparatively lowest model accuracy 68.69% and 48.09% for my dataset. The results demonstrate the effectiveness of CNNs in sensitive speech detection and classification tasks, showcasing their ability to accurately identify sensitive speech and non-sensitive speech. The research contributes to the development of robust tools and frameworks for content moderation in online platforms. Future work involves addressing these limitations and exploring advanced techniques, and multi-language attention mechanisms, to further improve the performance of sensitive speech detection systems. Ultimately, this research serves as a foundation for promoting respectful and inclusive digital communities by combating the detrimental impact of sensitive speech.