Emotion detection plays a crucial role in understanding human expressions and has widespread applications in social media analysis, conversational agents, and mental health monitoring. While significant advancements have been made for languages such as English and Bangla, the Sylheti dialect-particularly in its Banglish form, which blends Bengali and English in Roman script-remains largely underexplored. This study addresses this gap by developing a robust emotion detection system specifically tailored for Sylheti-Banglish text. A comprehensive dataset encompassing seven fundamental emotion classes-neutral, anger, joy, love, sadness, fear, and surprise-was meticulously curated from diverse digital platforms. The research tackles challenges inherent in unstructured, code-mixed, and dialectal text, including inconsistent grammar and mixed-script usage. To effectively model these complexities, three state-of-the-art transformer-based architectures were employed: Multilingual BERT (mBERT), XLM-RoBERTa, and BanglishBERT (ELECTRA). Each model was fine-tuned on the curated dataset, leveraging advanced preprocessing techniques such as normalization, tokenization, and oversampling to enhance performance. Experimental results demonstrate the effectiveness of transformer architectures in capturing the nuanced emotional expressions within Sylheti-Banglish text, with mBERT achieving particularly strong performance. These findings underscore the potential of multilingual and code-mixed language models in advancing emotion detection for underrepresented dialects and scripts. This research not only contributes a novel dataset and methodological framework but also lays a foundation for future studies in dialectal emotion analysis.