In recent years, with the proliferation of Internet platforms, each striving to enhance user engagement, corresponding social channels have been established to facilitate user expression of personal opinions and viewpoints. However, this openness has also provided a platform for some individuals to flout platform regulations and disseminate toxic content laden with violent tendencies. Such behavior undermines the stability of these platforms, jeopardizes national unity, and poses risks to societal cohesion and individual well-being. Recognizing and addressing the toxicity of online speech is paramount for enhancing user experience, fostering a healthier online environment, and advancing societal development. This paper introduces BLAM, an online speech toxicity detection model built upon deep learning principles. Leveraging TEDA training data on toxic speech instances, we propose a deep neural network model, BLAM, integrating bidirectional long short-term memory networks, self-attention mechanisms, and global maximum pooling layers. Our approach demonstrates an 8% improvement over traditional methods in accurately identifying toxic speech.