Bengali Speech Recognition Fixed by Tokenizer Swap

Researchers Sanjid Hasan and Md. Abdur Rahman have identified why compact speech recognition models fail catastrophically with Bengali and developed a surprisingly efficient fix. The problem lies not in the model architecture itself but in English-centric byte-level tokenizers that fragment Bengali words into unstable token chains, triggering what they call *autoregressive collapse* during inference.

Their solution, *vocabulary transplantation*, replaces the decoder vocabulary with BanglaBERT WordPiece tokens designed for Bengali's morphological richness and native script. This surgical intervention requires no costly pre-training from scratch. Results were dramatic: token fertility dropped from 9.16 to 1.30, and autoregressive sequence length fell by 85.8%.

Testing on the 882-hour Lipi-Ghor dataset yielded a 21.54% Word Error Rate and a Real-Time Factor of 0.0053, proving the approach enables both accurate and extremely fast edge deployment. Accepted at ICML 2026's MusIML Workshop, the research offers a scalable blueprint for adapting lightweight ASR systems to other morphologically complex, non-Latin languages historically underserved by English-centric AI infrastructure.