From Speech to Text Corpora: Evaluating ASR-Based Data Acquisition for Low-Resource Fongbe and Hausa

Researchers successfully fine-tuned automatic speech recognition (ASR) models to create text corpora for low-resource African languages Fongbe and Hausa, achieving significant improvements in transcription accuracy. The work demonstrates ASR's potential for rapidly expanding language resources in underrepresented languages, though quality varies by linguistic complexity, with Hausa transcriptions approaching production-ready standards while Fongbe requires further refinement.

This research addresses a critical bottleneck in natural language processing: the scarcity of training data for African languages. By leveraging existing ASR architectures (MMS-300M and Whisper-Small) and fine-tuning them on curated datasets, the researchers demonstrate a scalable pathway for bootstrapping language resources without waiting for manual transcription efforts. The 78% relative error reduction for Fongbe represents substantial technical progress, particularly given the language's tonal and diacritic complexity—challenges that typically plague ASR systems.

The work reflects broader trends in AI democratization, where transfer learning and pre-trained multilingual models enable resource-constrained teams to tackle previously intractable problems. Low-resource language NLP has traditionally lagged behind English and Mandarin, limiting downstream applications in machine translation, voice interfaces, and digital content accessibility for millions of speakers. This research contributes to closing that gap.

The human evaluation results—57.4/100 for Hausa versus 36.5/100 for Fongbe—reveal important truths about ASR performance across linguistic typologies. Non-tonal languages without heavy diacritical marking are more amenable to current ASR approaches, while tonal languages require either specialized architectures or substantial post-processing. The release of datasets and models following ethical guidelines enables community-driven improvements.

Looking ahead, developers should monitor whether similar ASR-bootstrapping approaches can achieve production quality for tonal African languages through hybrid human-in-the-loop workflows. The catalog of 1,553 YouTube videos represents a valuable resource for future work, potentially accelerating progress across multiple languages simultaneously.

• →Fine-tuned MMS-300M achieved 9.48% WER on Fongbe, a 78% improvement over previous baselines while preserving tonal diacritics.
• →ASR-generated corpora show promise for low-resource language NLP, with quality sufficient for production use in non-tonal languages like Hausa.
• →Tonal languages with diacritical marks require significantly more refinement than non-tonal languages when using current ASR pipelines.
• →Researchers curated 1,553 YouTube videos and processed 45.49 hours into 6,770 transcribed segments, establishing a scalable acquisition methodology.
• →Open release of datasets, models, and video catalog enables community-driven improvements for African language NLP.