In social networks, hate speech detection is an emerging need for avoiding offensive messages as well as to block the corresponding users. Many users are affected by these messages on social networks like Facebook, Twitter, LinkedIn, etc. Machine learning techniques have progressed significantly in detecting and classifying hate speech. Recently, the long-short term memory (LSTM) with term frequency inverse document frequency (TF-IDF) has been developed for hateful sentiment detection using tweets datasets. The LSTM with TF-IDF efficiently detected hate speech when compared with other successful existing classifier models, support vector machine (SVM), logistic regression (LR), random forest (RF), and bidirectional encoder representations from transformers (BERT) models. Hate speech detection becomes a more challenging issue than hateful sentiments. Thus, this paper focused on effective training of hate speech using either the extraction of TF-IDF or topic-based features. However, TF-IDF-based features become worse due to the occurrence of the scalability problem of large-sized tweet datasets. Thus, the proposed work is framed with two crucial steps to overcome this scalability issue, in which, initially, the topic-based features are extracted using the deep-based models since topic-based features take a smaller number of dimensions than TF-IDF-based features. These features are fused with classifier models to deliver efficient hate speech detection results. In the experimental work, a total of 24,783 labelled tweets were taken as hate, offensive language, or neither, and demonstrated that the proposed deep-based fusion model efficiently detects hate speech better than other classifiers.