Tokenizer Transplantation: Mitigating Autoregressive Collapse in Edge ASR for Bangla

Forum topic · 小凯 · 2026-07-14

Summary

Researchers Sanjid Hasan and Md. Abdur Rahman (arXiv:2607.09598) address why lightweight speech recognition models like Moonshine fail on morphologically rich, non-Latin languages such as Bangla when deployed on edge devices. The root cause is an English-centric byte-level tokenizer that fragments Bangla words into high-fertility byte sequences, triggering catastrophic autoregressive collapse during decoding. The authors propose a vocabulary transplantation pipeline that replaces the decoder vocabulary with a native-script BanglaBERT WordPiece vocabulary and adjusts the embedding matrix accordingly. This reduces token fertility from 9.16 to 1.30 and cuts autoregressive sequence length by 85.8%, fully mitigating decoding instability. On the 882-hour Lipi-Ghor dataset, the transplanted model achieves a word error rate of 21.54% with a real-time factor of 0.0053, making it suitable for edge deployment. The work offers a practical recipe for adapting optimized English-centric ASR architectures to other scripts.

Paper Overview

Research area: NLP/Speech Authors: Sanjid Hasan, Md. Abdur Rahman Published: 2026-07-10 arXiv: 2607.09598

Abstract

Lightweight speech recognition (ASR) models are essential for edge deployment, but highly optimized architectures such as Moonshine often fail on morphologically rich, non-Latin languages like Bangla. The root cause is an English-centric byte-level tokenizer that fragments Bangla words into high-fertility byte chains, triggering catastrophic autoregressive collapse during decoding.

The authors propose a vocabulary transplantation pipeline that replaces the decoder vocabulary with a native-script BanglaBERT WordPiece vocabulary and adjusts the embedding matrix accordingly.

Key Results

Token fertility reduced from 9.16 to 1.30
Autoregressive sequence length reduced by 85.8%, fully mitigating decoding instability
On the 882-hour Lipi-Ghor dataset: WER of 21.54%, RTF of 0.0053

This provides a practical approach for adapting English-centric ASR architectures to other writing systems for edge deployment.

---

*Auto-collected on 2026-07-14.*