Regional dialects of Bangla, such as Sylheti and Chittagonian, pose significant challenges for natural language processing due to their low-resource nature and substantial linguistic variation from standard Bangla. In this work, we present a systematic evaluation of eight open-source LLMs for translating fifteen distinct Bangla dialects into standard Bangla. To achieve this comprehensive coverage, we utilize a combination of established benchmarks and a novel dataset curated from an ongoing regional linguistic project. We assess model performance using a multi-metric framework that combines exact-match and error-rate evaluations such as, Averaged BLEU, WER, and CER with embedding-based semantic metrics including BERTScore, METEOR, and COMET. Additionally, we perform a detailed dialect-level linguistic analysis to identify the deep-seated structural, orthographic, and semantic barriers inherent to dialectal translation. Our study highlights the strengths and limitations of current open-source models, provides empirical insights for future dialect-aware fine-tuning, and contributes a reproducible benchmark for the research community.

Conventional research on speech recognition modeling relies on the canonical form for most low-resource languages while automatic speech recognition (ASR) for regional dialects is treated as a fine-tuning task. To investigate the effects of dialectal variations on ASR we develop a 78-hour annotated Bengali Speech-to-Text (STT) corpus named Ben-10. Investigation from linguistic and data-driven perspectives shows that speech foundation models struggle heavily in regional dialect ASR, both in zero-shot and fine-tuned settings. We observe that all deep learning methods struggle to model speech data under dialectal variations, but dialect specific model training alleviates the issue. Our dataset also serves as a out-of-distribution (OOD) resource for ASR modeling under constrained resources in ASR algorithms. The dataset and code developed for this project are publicly available.