Selected work

Representative Papers

Tone-Conditioned Curriculum Learning for Low-Resource Bantu Speech Recognition

Jun 30, 2026

Existing foundation models exhibit zero-shot word error rates (WER) exceeding 100% on Southern Bantu speech recognition tasks, rendering them impractical. This work proposes a tone-conditioned curriculum learning framework that, for the first time, integrates tonal information into the curriculum mechanism. By combining mixed-difficulty scoring, tone-statistics-gated adapters, and a staged training strategy, the approach significantly enhances cross-lingual transfer performance for low-resource Bantu languages. Experiments on W2V-BERT and Whisper demonstrate that the method reduces average WER to 28.41% across six languages, achieving 23.79% on the Xitsonga transfer task. The study further uncovers systematic interactions between model architecture and language family, offering an effective adaptive training paradigm for low-resource speech recognition.

0 citationsRead paper

AfriScience-MT: Towards Decolonizing Science in Africa through Text Translation

May 28, 2026

This study addresses the marginalization of African languages in scientific communication, which impedes access to and production of scientific knowledge for hundreds of millions of speakers due to the absence of standardized scientific terminology. To bridge this gap, the authors present AfriScience-MT, the first systematically constructed parallel corpus spanning six African languages and eleven scientific domains, developed through collaboration between professional translators and science communication experts who translated paper abstracts and coined missing technical terms. The resource enables machine translation evaluation under zero-shot, few-shot, and fine-tuned settings. Evaluations reveal that closed-source models (GPT-5.4 and Gemini) achieve state-of-the-art performance (sentence-level COMET scores of 68.3 and 68.0, respectively), while fine-tuned open-source models (NLLB-1.3B and TranslateGemma-12B) also demonstrate strong results. This work advances the decolonization and localization of scientific knowledge and fills a critical void in African-language scientific translation resources.

0 citationsRead paper

The Annotation Scarcity Paradox in Low-Resource NLP Evaluation: A Decade of Acceleration and Emerging Constraints

May 18, 2026

This study addresses a structural tension in low-resource natural language processing (NLP): the rapid expansion of technical capabilities juxtaposed with a scarcity of linguistic expertise needed for robust evaluation, giving rise to the “annotation scarcity paradox”—where advances in model performance exacerbate evaluative inequities. Through a critical narrative review spanning 2014 to 2026, the paper systematically examines validity threats and power imbalances inherent in current evaluation practices, positioning evaluation sovereignty and knowledge governance at the core of methodological reform. It advocates shifting from extractive, transactional data paradigms toward community-embedded, relational evaluation frameworks. To this end, the work integrates techniques such as data augmentation, model-based evaluation, participatory curation, item response theory, and active learning to construct a next-generation evaluation framework for low-resource NLP that jointly upholds fairness, validity, and data sovereignty.

0 citationsRead paper

Calibrating Beyond English: Language Diversity for Better Quantized Multilingual LLM

Jan 26, 2026

This study addresses the performance degradation of multilingual large language models under post-training quantization when calibrated exclusively with English data. The authors systematically evaluate five monolingual and three multilingual mixed calibration strategies across ten languages using both GPTQ and AWQ quantization methods, with experiments conducted on Llama3.1-8B and Qwen2.5-7B. Their findings reveal, for the first time, that incorporating multilingual or non-English calibration data significantly improves post-quantization perplexity. Notably, language-aligned calibration strategies consistently outperform generic English-only calibration, reducing average perplexity by up to 3.52 points. These results underscore the critical role of linguistic diversity in the calibration process for effective multilingual model quantization.

0 citationsRead paper

Swivuriso: The South African Next Voices Multilingual Speech Dataset

Dec 01, 2025

High-quality, multi-domain speech data for automatic speech recognition (ASR) in South Africa’s seven official languages is critically scarce, hindering robust ASR development for these low-resource languages. Method: We construct the first large-scale, multi-domain speech dataset covering all seven languages—totaling 3,000 hours—with content spanning agriculture, healthcare, and other domains. Data collection adheres to strict ethical guidelines; transcripts undergo rigorous human verification, and audio is processed via standardized preprocessing to ensure high fidelity and annotation accuracy. Contribution/Results: This dataset fills a critical gap in benchmark resources for indigenous low-resource ASR in South Africa, enabling end-to-end model training and cross-domain acoustic modeling. Extensive experiments across mainstream ASR architectures—including conformer, wav2vec 2.0, and Whisper—demonstrate that models trained or fine-tuned on our dataset consistently outperform those based on existing public benchmarks. The dataset provides a reproducible, scalable foundation for evaluation and development of ASR systems for under-resourced languages.

0 citationsRead paper

Recent publications

Latest Papers

Tone-Conditioned Curriculum Learning for Low-Resource Bantu Speech Recognition

Jun 30, 2026

Existing foundation models exhibit zero-shot word error rates (WER) exceeding 100% on Southern Bantu speech recognition tasks, rendering them impractical. This work proposes a tone-conditioned curriculum learning framework that, for the first time, integrates tonal information into the curriculum mechanism. By combining mixed-difficulty scoring, tone-statistics-gated adapters, and a staged training strategy, the approach significantly enhances cross-lingual transfer performance for low-resource Bantu languages. Experiments on W2V-BERT and Whisper demonstrate that the method reduces average WER to 28.41% across six languages, achieving 23.79% on the Xitsonga transfer task. The study further uncovers systematic interactions between model architecture and language family, offering an effective adaptive training paradigm for low-resource speech recognition.

0 citationsRead paper

AfriScience-MT: Towards Decolonizing Science in Africa through Text Translation

May 28, 2026

This study addresses the marginalization of African languages in scientific communication, which impedes access to and production of scientific knowledge for hundreds of millions of speakers due to the absence of standardized scientific terminology. To bridge this gap, the authors present AfriScience-MT, the first systematically constructed parallel corpus spanning six African languages and eleven scientific domains, developed through collaboration between professional translators and science communication experts who translated paper abstracts and coined missing technical terms. The resource enables machine translation evaluation under zero-shot, few-shot, and fine-tuned settings. Evaluations reveal that closed-source models (GPT-5.4 and Gemini) achieve state-of-the-art performance (sentence-level COMET scores of 68.3 and 68.0, respectively), while fine-tuned open-source models (NLLB-1.3B and TranslateGemma-12B) also demonstrate strong results. This work advances the decolonization and localization of scientific knowledge and fills a critical void in African-language scientific translation resources.

0 citationsRead paper

The Annotation Scarcity Paradox in Low-Resource NLP Evaluation: A Decade of Acceleration and Emerging Constraints

May 18, 2026

This study addresses a structural tension in low-resource natural language processing (NLP): the rapid expansion of technical capabilities juxtaposed with a scarcity of linguistic expertise needed for robust evaluation, giving rise to the “annotation scarcity paradox”—where advances in model performance exacerbate evaluative inequities. Through a critical narrative review spanning 2014 to 2026, the paper systematically examines validity threats and power imbalances inherent in current evaluation practices, positioning evaluation sovereignty and knowledge governance at the core of methodological reform. It advocates shifting from extractive, transactional data paradigms toward community-embedded, relational evaluation frameworks. To this end, the work integrates techniques such as data augmentation, model-based evaluation, participatory curation, item response theory, and active learning to construct a next-generation evaluation framework for low-resource NLP that jointly upholds fairness, validity, and data sovereignty.

0 citationsRead paper

Calibrating Beyond English: Language Diversity for Better Quantized Multilingual LLM

Jan 26, 2026

This study addresses the performance degradation of multilingual large language models under post-training quantization when calibrated exclusively with English data. The authors systematically evaluate five monolingual and three multilingual mixed calibration strategies across ten languages using both GPTQ and AWQ quantization methods, with experiments conducted on Llama3.1-8B and Qwen2.5-7B. Their findings reveal, for the first time, that incorporating multilingual or non-English calibration data significantly improves post-quantization perplexity. Notably, language-aligned calibration strategies consistently outperform generic English-only calibration, reducing average perplexity by up to 3.52 points. These results underscore the critical role of linguistic diversity in the calibration process for effective multilingual model quantization.

0 citationsRead paper

Swivuriso: The South African Next Voices Multilingual Speech Dataset

Dec 01, 2025

High-quality, multi-domain speech data for automatic speech recognition (ASR) in South Africa’s seven official languages is critically scarce, hindering robust ASR development for these low-resource languages. Method: We construct the first large-scale, multi-domain speech dataset covering all seven languages—totaling 3,000 hours—with content spanning agriculture, healthcare, and other domains. Data collection adheres to strict ethical guidelines; transcripts undergo rigorous human verification, and audio is processed via standardized preprocessing to ensure high fidelity and annotation accuracy. Contribution/Results: This dataset fills a critical gap in benchmark resources for indigenous low-resource ASR in South Africa, enabling end-to-end model training and cross-domain acoustic modeling. Extensive experiments across mainstream ASR architectures—including conformer, wav2vec 2.0, and Whisper—demonstrate that models trained or fine-tuned on our dataset consistently outperform those based on existing public benchmarks. The dataset provides a reproducible, scalable foundation for evaluation and development of ASR systems for under-resourced languages.

0 citationsRead paper