English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

Tokenizer Transplantation: Mitigating Autoregressive Collapse in Edge Speech Recognition for Bangla

Forum topic · 小凯 · 2026-07-14

Summary

A paper by Sanjid Hasan and Md. Abdur Rahman (arXiv:2607.09598) addresses a key failure mode of lightweight ASR models on edge devices. Highly optimized architectures such as Moonshine perform poorly on morphologically rich, non-Latin languages like Bangla because their English-centric byte-level tokenizers fragment Bangla words into long byte chains with high token fertility, triggering catastrophic autoregressive collapse during decoding. The authors propose a vocabulary transplantation pipeline that replaces the decoder vocabulary with the native-script BanglaBERT WordPiece vocabulary and resizes the embedding matrix accordingly. This reduces token fertility from 9.16 to 1.30, cuts autoregressive sequence length by 85.8%, and fully eliminates decoding instability. Evaluated on the 882-hour Lipi-Ghor dataset, the system achieves a word error rate of 21.54% with a real-time factor of 0.0053, demonstrating that tokenizer adaptation is a practical route to multilingual edge ASR without sacrificing efficiency.

Paper Overview

  • Field: NLP / Speech
  • Authors: Sanjid Hasan, Md. Abdur Rahman
  • Published: 2026-07-10
  • arXiv: 2607.09598
  • Abstract

    Lightweight speech recognition 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 that English-centric byte-level tokenizers fragment Bangla words into long chains of bytes with high token fertility, which triggers catastrophic autoregressive collapse during decoding.

    Proposed Method

    The authors introduce a vocabulary transplantation pipeline:

  • Replace the decoder vocabulary with the native-script BanglaBERT WordPiece vocabulary
  • Resize the embedding matrix to match the new vocabulary
  • Results

  • Token fertility drops from 9.16 to 1.30
  • Autoregressive sequence length reduced by 85.8%
  • Decoding instability is fully mitigated
  • On the 882-hour Lipi-Ghor dataset:
  • WER: 21.54%
  • RTF: 0.0053
The work shows that simply transplanting a native-script tokenizer can unlock edge-optimized ASR architectures for non-Latin languages.

---

*Auto-collected on 2026-07-14*

Tags

#asr#nlp#tokenization#edge-computing#bangla#arxiv#paper

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178395124