Researchers released MameLoshnLM, described as the first open-source 8-billion-parameter language model dedicated to Yiddish. The work also includes an evaluation benchmark intended to fill the gap in reliable testing resources for the language. It addresses the low digital availability of Yiddish text despite its substantial written heritage.
Researchers present Enemray, a language model designed for general-purpose interaction in Hassaniya, a low-resource Arabic variety. Training follows a stability-plasticity objective intended to build strong linguistic and cultural competence while limiting degradation of the model's existing capabilities. The work is published as an arXiv preprint.
A new arXiv paper proposes an online adaptive sampling strategy for realigning multilingual language models, aiming to improve cross-lingual transfer to extremely low-resource languages. The authors note that existing realignment approaches typically use uniform or random sampling, which may underuse informative language pairs. Their method adjusts sampling dynamically as training proceeds within a distributed setup.
A new arXiv study systematically examines token merging as a way to cut the computational cost of large multilingual speech recognition models such as Whisper. The technique dynamically combines token representations during inference, and the authors test how its effectiveness varies with model size and fine-tuning. The work targets deployment efficiency for transcribing low-resource languages without language-specific training.
A new arXiv preprint describes a method that stacks adapters sequentially to help large multilingual speech recognition models handle low-resource languages. The authors note that current systems perform unevenly, favoring high-resource languages and losing accuracy where labeled audio is scarce. The abstract covers the approach at a high level, with results and evaluation details not included in the excerpt.
Researchers added Portuguese-source evaluation data for Xichangana, Mozambican Nyanja, and Sena to the FLORES+ multilingual benchmark. The work examines how conflating closely related language varieties affects machine translation evaluation, comparing Xichangana against Tsonga and Mozambican Nyanja against Chichewa. It argues that merging distinct varieties into a single reference can distort measured translation quality.
Researchers present a new annotated corpus for classifying sentence functions in Bangla, a resource previously lacking for the language. The work benchmarks several models on the task and adds interpretability analysis of their predictions. Such sentence-type identification supports dialogue systems, speech synthesis, and machine translation.
A new arXiv preprint argues that word error rate alone hides important failures when speech recognition systems and audio language models handle code-switched speech. The authors propose an evaluation method that accounts for language switches, and apply it to English-Yoruba audio, a low-resource pair with diacritics. They find that strong monolingual benchmark scores do not carry over to this setting.
A new arXiv paper investigates how well multilingual large language models handle open-ended text generation in Urdu, a low-resource language. The authors argue that models marketed as multilingual often fall short in cultural and linguistic correctness outside high-resource languages. The work contributes to broader questions about the reliability of these systems for non-English users.
A new arXiv paper introduces E-CONAN, a set of benchmarks covering entailment, contradiction and neutral relations for Arabic natural language inference. The authors frame the work as a response to the limited resources available for Arabic compared with English and other well-served languages, noting that inference models are a component of many downstream NLP applications. The datasets are intended to support training and evaluation of Arabic inference systems.
A revised arXiv paper examines how large language models can be adapted to detect hate speech in Roman Urdu, a low-resource language written in Latin script. The authors compare several adaptation approaches, addressing challenges such as scarce annotated data, informal writing conventions, and the lack of standardized grammar. The work focuses on efficient methods suited to settings where labeled corpora are limited.
A new arXiv paper introduces SALT, a technique for strengthening how individual tokens are represented within multilingual sentence encoders. These encoders are optimized to align whole sentences across many languages, supporting applications like translation mining and zero-shot learning for low-resource languages. The paper targets the weaker token-level alignment that results from this sentence-focused training.
A new arXiv paper introduces 5-Dialects-BN, a resource covering five Bangla dialects used to examine how transliteration choices influence large language model performance. The work targets the sharp accuracy drops LLMs show on low-resource, dialectally diverse languages, with Bangla being the world's sixth most spoken language. The authors position script and transliteration practices as an underexamined factor in this degradation.
Researchers have released BaltiVoice, a 16.8-hour read-speech corpus with 10,060 validated utterances for Balti, a Tibetic language spoken in Gilgit-Baltistan, Pakistan. The language previously had no publicly available speech recognition resources, making this the first open dataset and ASR model for Balti. The team fine-tuned OpenAI's Whisper architecture on the corpus to enable automatic speech recognition for the language.
Researchers present a data curation recipe combined with deterministic prompting to keep speaker identity stable in text-to-speech systems trained on limited speech data. Modern Greek serves as the test case, since it lacks the curated corpora that underpin state-of-the-art synthesis for high-resource languages. The work addresses quality degradation that TTS models typically show when clean training speech is scarce.
A new arXiv paper introduces BuzzASR, a collection of more than one hundred monolingual Whisper models fine-tuned for automatic speech recognition in 102 languages. The language-specialized models are designed to cover languages that large general-purpose multilingual ASR systems often serve poorly. The release is presented as a resource for practitioners and researchers working on speech technology across many languages.
Hugging Face's Open ASR Leaderboard has expanded its coverage to include a language from the Global South for the first time. The addition broadens the benchmark's evaluation of automatic speech recognition systems beyond the predominantly high-resource languages it previously tracked. It reflects a wider push to measure model performance on underrepresented languages.
WHY IT MATTERS ↘Benchmarks drive where engineering effort goes, so extending a widely cited ASR leaderboard to a Global South language gives vendors and researchers a shared target for measuring quality on languages that commercial incentives alone have largely ignored. The caveat is that a single added language still reflects an underrepresentative sample, so teams should treat it as a starting signal for data collection and evaluation rather than evidence of broad multilingual coverage.