Tokenizer Transplantation: Mitigating Autoregressive Collapse in Edge-Efficient Bengali ASR
Quick Answer
This study introduces a vocabulary transplantation pipeline to enhance Bengali ASR performance by replacing the English-centric tokenizer with BanglaBERT WordPiece vocabulary, achieving a 21.54% Word Error Rate and reducing autoregressive sequence length by 85.8%.
Quick Take
This study introduces a vocabulary transplantation pipeline to enhance Bengali ASR performance by replacing the English-centric tokenizer with BanglaBERT WordPiece vocabulary, achieving a 21.54% Word Error Rate and reducing autoregressive sequence length by 85.8%. The approach mitigates decoding instability, making it a scalable solution for edge deployment in morphologically rich languages.
Key Points
- Novel pipeline replaces English tokenizer with BanglaBERT vocabulary for Bengali ASR.
- Achieved 21.54% Word Error Rate on the 882-hour Lipi-Ghor dataset.
- Reduced token fertility from 9.16 to 1.30, improving model stability.
- Decreased autoregressive sequence length by 85.8%, eliminating decoding instability.
- Provides a reproducible blueprint for cross-script adaptation of ASR models.
Paper Resources
📖 Reader Mode
~2 min readAbstract:Lightweight speech recognition models are critical for edge deployment, yet highly optimized architectures like Moonshine often fail on morphologically rich, non-Latin languages such as Bengali. This study identifies the root cause of this failure as the model's English-centric byte-level tokenizer, which fragments Bengali words into high-fertility byte chains and triggers catastrophic autoregressive collapse during inference. To resolve this, a novel vocabulary transplantation pipeline is proposed to replace the decoder vocabulary with the native-script BanglaBERT WordPiece vocabulary and resize the corresponding token embedding matrix. Experimental results demonstrate a reduction in token fertility from 9.16 to 1.30. By decreasing autoregressive sequence length by 85.8%, decoding instability is entirely mitigated. When evaluated on the 882-hour Lipi-Ghor dataset, the modified architecture achieves a competitive 21.54% Word Error Rate (WER) and a Real-Time Factor (RTF) of 0.0053. Ultimately, this research provides a scalable, reproducible blueprint for cross-script adaptation of compact ASR models without the need for resource-intensive pre-training.
| Comments: | 5 pages, 2 figures. Accepted as a poster at the MusIML Workshop, ICML 2026 |
| Subjects: | Computation and Language (cs.CL) |
| ACM classes: | I.2.7; I.2.6 |
| Cite as: | arXiv:2607.09598 [cs.CL] |
| (or arXiv:2607.09598v1 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2607.09598 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Sanjid Hasan [view email]
[v1]
Fri, 10 Jul 2026 16:54:38 UTC (1,030 KB)
— Originally published at arxiv.org
Want this in your inbox every morning?
Daily brief at your local 8am — bilingual EN/中文, free.
More from arXiv cs.CL
See more →Letting the Data Speak: Extracting Keywords from Crowdsourced Collections with AI
The study evaluates three NLP approaches—Named Entity Recognition, Keyword Extraction, and Topic Modelling—using the Their Finest Hour Online Archive to automate keyword extraction from crowdsourced WWII collections. Findings suggest that while NLP methods show promise, no single approach is sufficient, and ethical considerations in automated keyword extraction are crucial for responsible stewardship.