Aspect Based Offensive Speech Detection Exploring Audio Features Through NLP Model

Emotion recognition system is a very common yet challenging area where research is still extending its exploration. The problem becomes more challenging when it comes to analyze the sentiments to be understandable by the autistic children from speech. This challenge has motivated us to implement a sentiment analysis system along with that the system will also detect the system will detect offensive content from the speech data. The analysis of sentiment by using the concept of Natural Language Processing is identifying and categorising emotions in speech data. The identification of speech that contains words which impacts a person or a group is known as offensive speech. The proposed task is focussed on identifying emotion in a person voice. After the identification of speech others task is on determining the nature of the content of the speech data. This has been achieved by computing the speech polarities. The category of the content is majorly categorising into positive, negative, or neutral. The primary focus of the propound research is on identifying inappropriate vocabulary as well sentiments majorly categories as stressful, happy, anger and disgust. The speech data we have taken is a monolingual audio speech data. The proposed method has adopted the concept of transformer in natural language processing approach for creating the data into vectors and uses deep neural network for categorizing offensive content from the speech data. The proposed work has extracted relevant speech-based audio features to find the emotion and applying deep neural network to train the speech data in order to find the abusive word has been detected by applying natural language processing approach. The model has been trained on well-known Natural language processing model such as Global Vectors for Word Representation (GloVe), Convolution Neural Network (CNN), Long Short-Term Memory (LSTM), Bidirectional Encoder Representations from Transformers (BERT) and RoBERTa. In order to make the proposed worked more robust, it has been trained standard benchmark data of audio speech along with that, it has been trained on sample data. Finally, the model has been compared with state of art work. Some statistical measure is also considered for better analysis of the proposed methodology.