<p>Incidences of hate speech, which could pose a threat to digital well being and free exchange of opinions, have heightened because of the recent blistering growth of content created by the users of the social media networks like Twitter, Facebook and Instagram. Whereas the traditional models emphasize binary classification of toxic content, they are mostly limited to be precise in reflecting the severity and dimensions of the various types of hate speech. The proposed research will introduce a new architectural ensemble based on workflows of Distil BERT to achieve efficient state of the art contextual embedding and a Bidirectional Long Short Term Memory (Bi-LSTM) network to capture sequential dependencies. Bayesian optimization, which is applied to optimize hyperparameters, is one method of enhancing the performance of a model. The given model was applied to the Jigsaw Toxic Comment Classification data and achieved higher results and accuracy of 95, precision of 93, recall of 95, and F1 score of 94. It should be noted that it also improves detection of underrepresented and classes for instance toxic, obscene, insult, severe toxic, threat, and identity remarkably. The outcomes show the usefulness of the model being implemented as a multi class hate speech levels detecting mechanism that can also form a viable content moderation system on the real time basis and provide the digital space with a safer experience.</p>