The creation and spread of misinformation through social media in regional languages is an escalating problem. There is a need to scrutinize these fabricated multimedia contents, especially in a low-resource language context. Current limitations in fake news detection research underscore the need for a comprehensive multimodal dataset in the Tamil language. This study proposes a multimodal dataset for fake news classification in Tamil, comprising 7,934 data samples across three modalities: text, image, and speech, curated from various Tamil fact-check websites. We develop the first Tamil speech corpus for fake news classification, encompassing 884 min of audio derived from fact-checked news articles, facilitating multimodal analysis. Additionally, we have gathered a substantial collection of 7934 images from fact-check Tamil news articles to explore the relationship between visual and textual data in the context of verified news. The dataset comprises thirteen attributes (text, media, and metadata). Furthermore, we evaluate the performance of baseline models for news classification using three multilingual BERT models for text with wav2vec2-large-xlsr-53 model for speech and Vision Transformer model for image along with four machine learning classifiers. Among all models, the logistic regression classifier trained on IndicBERT feature embeddings has achieved the highest F1 score of 0.6701, a recall of 0.6931 and a precision of 0.6585 along with an accuracy of 69.31%. Our proposed strategy for creating multimodal datasets can be extended to other low-resource languages, advancing the field of fake news classification research.