Bangla is a diglossic and morphologically rich language with two primary registers—Sadhu Bhasha (classical) and Cholit Bhasha (colloquial)—and identifying these registers is crucial for downstream NLP tasks such as translation, OCR, and speech synthesis. This study developed a balanced dataset of 7,350 sentences from both registers, applied preprocessing steps including tokenization, normalization, and padding, and performed an 80−20 train–test split. Four deep learning models—LSTM, Bi-LSTM, BanglaBERT, and mBERT—were trained under identical settings using the Adam optimizer with a batch size of 32 for 10 epochs. While the sequential models showed reasonable performance, transformer-based architectures outperformed them substantially, with BanglaBERT achieving the highest accuracy of 95%, establishing a strong benchmark for Sadhu–Cholit classification and emphasizing the importance of register sensitivity in Bangla NLP.