Detecting hate content in social media poses significant challenges due to limited annotated speech data, optimal model selection, and feature selection methods. This challenge is particularly pronounced for Dravidian languages, which lack sufficient datasets. To tackle this issue, we introduce the multimodal dataset titled DravLangGuard for hate content detection in Dravidian social media. Sourced from YouTube, the dataset includes 921 non-hate and 1,172 hate speech utterances, along with annotated text for three Dravidian languages. Furthermore, the dataset categorizes hate speech into gender, religious, personal defamation, and political/nationality affiliation categories. We developed a baseline multimodal hate speech recognition system for this created dataset using Mel spectrogram and Language-agnostic BERT Sentence Embedding sentence embeddings with a 1D-Convolutional neural network as the backbone. Additionally, ablation studies based on modality were conducted, demonstrating superior performance of the proposed model. The multimodal approach attained macro F1 scores of 0.8078 for Malayalam, 0.7105 for Tamil, and 0.7524 for Telugu.