Food security depends on correct crop disease identification; nevertheless, linguistic obstacles frequently prevent farmers from utilizing automated diagnostic devices. For farmers who use their local Bengali to describe crop diseases, this article addresses this problem. With the help of a fresh, specifically constructed dataset of 4,500 entries, we created and tested a system for categorizing diseases based on these statements. Training and testing computer models rely on this resource, which is balanced throughout 15 disease categories to offer a solid base. We compared deep neural network models and conventional machine learning against cutting-edge Transformers in our thorough investigation. The findings supported Transformers’ better performance; the language-specific Bangla-BERT had the highest accuracy at 95.04%. Other multilingual models like Distil-BERT, m-BERT, and XLM-R also outperformed all non-Transformer architectures by a wide margin, which shows how important contextual language understanding is for this job. The primary results of this research are the validation of the efficacy of language-specific models for practical agricultural applications and the availability of a helpful public dataset. These results clear the path for the creation of more fair and easily available diagnostic instruments for farmers in their own tongue.