Tokenizer transplantation mitigates autoregressive collapse in edge-efficient Bengali ASR

Researchers address the failure of lightweight speech recognition models like Moonshine on morphologically rich languages such as Bengali by replacing their English-centric byte-level tokenizers with native-script BanglaBERT WordPiece vocabularies. This modification resizes the token embedding matrix and significantly reduces token fertility from 9.16 to 1.30.

• The approach decreases autoregressive sequence length by 85.8%, entirely mitigating decoding instability.
• Evaluated on the 882-hour Lipi-Ghor dataset, the modified architecture achieves a Word Error Rate (WER) of 21.54% and a Real-Time Factor (RTF) of 0.0053.

This research provides a scalable, reproducible blueprint for cross-script adaptation of compact ASR models without requiring resource-intensive pre-training.