An Integrated Embedded System Towards Abusive Bengali Speech and Speaker Detection Using NLP and Deep Learning

Intelligible speech, while it provides an excellent means of communication for humans and sets us apart from other lifeforms, our abuse of speech creates deep and lasting issues in our society. The use of derogatory language has a significant impact not only on children’s mental health but also on adults, for instance, in an abusive work environment. Accountability for such actions is one of the key steps toward maintaining a healthy atmosphere or at least making it less frequent. In this paper, we describe our work on detecting abusive or hate speech in Bangla in real time. Our system converts the speech to text and then uses NLP and deep learning to detect such occurrences in real-time. Also, if the voice is registered on our system, it identifies the person engaging in abusive words, opening ways to greater workplace accountability. We also describe our mobile application and the microcontroller-based standalone embedded system that can be deployed in target places (for instance, daycare centers, schools, workplaces, etc.) to record audio and detect the abusive speech and the speaker in real-time. Several datasets have been deployed on the LSTM, Bi-LSTM, GRU, and BERT models to assess the system’s efficacy. Identification of the individual speaking the words is done using the audio signal extraction feature MFCC. The experimental results show that the BERT model provides the highest accuracy compared to other algorithms.