cs.CLNov 12, 2026

Low Language Transfer Speech-to-Text

Authors: Husein Zolkepli

Organizations: Scicom (MSC) Berhad, Malaysia

Abstract

A multilingual speech recogniser has heard much speech in its own languages. We ask how much of that it brings to a language it was never trained on, when it is fine-tuned on the little that language has. We score 28 open multilingual models untouched on Low-Language-TTS, 50 low-resource languages in 11 writing systems, and fine-tune 24 of them on at most 1,000 clips a language, 82 hours in all, with a dev set choosing every epoch. Fine-tuning lowers every model's macro character error rate (CER), by 16.6 to 53.1 points, to 9.5 at best, and a word language model takes three CTC models 1.2 to 1.4 points lower; WER is still 34.4 at best. On the 12 languages that no model lists, every model starts at 46 or worse and the best ends at 11.7. The same architectures from random weights stay 60 points or more behind, so the gain is transfer, and an encoder left untouched already gives a Whisper or Qwen decoder most of what it needs. It is the language, not the corpus: on FLEURS recordings that ours do not share, the median model keeps 79% of its gain. It costs the languages the models knew, 5.7 points of CER on English, Spanish, French and German in the median model. Low-rank adapters on 4% of the weights come within 1.1 points of tuning everything, and on 2 of 3 models forget less. Seq2seq and language-model decoders are done within three to five epochs, while a new CTC head is still improving at twenty or thirty. Data beats epochs: 25 clips a language take Qwen3-ASR 0.6B from 60.9 to 39.3, and at equal clips presented, ten times the data is worth 9.7 points. A model that first learned forty of the languages learns the other ten no faster. On a public Malay benchmark, untouched Whisper large-v3 scores 8.9 when it detects the language and 14.9 when told the clip is Malay, because it then translates the English that speakers mix in; trained on the Scicom telephony dataset it scores 6.0.

Figures & tables

Appendix figures & tables1 asset

Supplementary material from the paper’s appendix.

Appendix

Explore similar work

Sep 14, 2026

Sequential Adapter Stacking for Cross-Lingual Low-Resource ASR

Extending large-scale multilingual automatic speech recognition (ASR) models to low-resource languages remains challenging. Model performance is skewed toward high-resource languages and degrades sharply for languages with limited labeled data and pre-training exposure. To address this, we investigate parameter-efficient approaches for transferring knowledge from resource-rich source languages to low-resource target languages on Whisper. Alongside warm initialization and attention-based fusion, we propose Sequential Adapter Stacking, which places a trainable target-language adapter on top of a frozen source-language adapter. Under controlled experiments, these approaches are evaluated on three target languages unsupported by Whisper -- Asturian, Assamese, and Xhosa -- using source languages with varying degrees of relatedness. Sequential Adapter Stacking with the closest related source consistently and significantly outperforms full fine-tuning across the three targets, with 5--8% relative WER reductions. These gains largely persist with only one hour of target training data.
Oct 28, 2026

How Architecture, Typology, and Adaptation Strategies Shape ASR for Low-Resource African Languages: A Controlled Empirical Study

Despite recent advancements in ASR, identifying effective architectures, fine-tuning approaches, and data adaptation strategies for low-resource languages remains challenging. We present a systematic comparison across five low-resource African languages and Swahili as a mid-resource contrast, fine-tuning three models from distinct architectures: Phi-4-multimodal-instruct, Whisper-large-v3-turbo, and MMS-1B-all. We analyze data sufficiency, monolingual versus multilingual adaptation, domain fine-tuning, and language-adapter transfer. Phi-4 achieves the best overall accuracy, MMS shows comparable CER, while Whisper exhibits high CER, consistent with known hallucinations and repetition loops. Performance saturates at different data scales, offering practical data-efficiency targets, while typological complexity consistently shapes error rates. Domain fine-tuning benefits Phi-4 and Whisper but degrades MMS, revealing limitations of adapter-based adaptation. Proxy language adapters outperform training MMS adapters from scratch.
Mar 22, 2026

Left Behind: Cross-Lingual Transfer as a Bridge for Low-Resource Languages in Large Language Models

We investigate how large language models perform on low-resource languages by benchmarking eight LLMs across five experimental conditions in English, Kazakh, and Mongolian. Using 50 hand-crafted questions spanning factual, reasoning, technical, and culturally grounded categories, we evaluate 2,000 responses on accuracy, fluency, and completeness. We find a consistent performance gap of 13.8-16.7 percentage points between English and low-resource language conditions, with models maintaining surface-level fluency while producing significantly less accurate content. Cross-lingual transfer-prompting models to reason in English before translating back-yields selective gains for bilingual architectures (+2.2pp to +4.3pp) but provides no benefit to English-dominant models. Our results demonstrate that current LLMs systematically underserve low-resource language communities, and that effective mitigation strategies are architecture-dependent rather than universal.