Researchers Sanjid Hasan and Md. Abdur Rahman identified a fundamental flaw in compact speech recognition models built for edge devices when applied to Bengali, a morphologically rich language with its own script. The lightweight Moonshine architecture, optimized for English, suffers from catastrophic autoregressive collapse during inference due to its English-centric byte-level tokenizer, which fragments Bengali words into long chains of high-fertility tokens. Rather than retraining the entire model, the researchers propose a method called vocabulary transplantation, which replaces Moonshine’s original decoder vocabulary with the BanglaBERT WordPiece vocabulary—a native-script vocabulary designed specifically for Bengali. This targeted surgical replacement avoids resource-intensive pre-training and addresses the structural decoding failure.