Social media networking sites have crossed the boundaries of traditional communication methods. English has traditionally been the primary language for comments and posts on social media. However, in recent years, a significant increase has been observed in using code-mixing language. The ease of communicating in the native language has enabled the user to create, publish, and share content that is a mix of native language + English language or native language written in English. The downside of using code-mixed language has led to the increased use of hateful and offensive words in social media posts. Pretrained word embeddings are crucial in extracting features for detecting hate speech using machine learning (ML) models. This paper explores the effectiveness of dual pretrained word embeddings: FastText and GloVe combined with TextCNN and Bi-GRU models for detecting hate speech and offensive language in English on social media platforms. The experiments demonstrate that dual word embeddings significantly improve hate speech detection accuracy compared to single pretrained word embeddings. The feature vectors obtained by integrating TF-IDF-weighted character n-grams and word trigram features improve the ML algorithm accuracy for code-mixed Hinglish hate speech content. Pretrained multilingual BERT embeddings enhance the model’s precision for the Hindi-English code-mixed hate speech dataset.