From Speech to Text Corpora: Evaluating ASR-Based Data Acquisition for Low-Resource Fongbe and Hausa
📰 ArXiv cs.AI
Learn how to use ASR pipelines to acquire text data for low-resource languages like Fongbe and Hausa, and evaluate their effectiveness
Action Steps
- Fine-tune a pre-trained ASR model like MMS-300M on a curated dataset for a low-resource language
- Evaluate the performance of the ASR model using metrics like WER on a benchmark like ALFFA
- Compare the results with a baseline model to measure the relative improvement
- Apply the ASR pipeline to extend text resources for other low-resource languages
- Test the quality of the acquired text data and its impact on language model training
Who Needs to Know This
NLP engineers and researchers working on low-resource languages can benefit from this approach to acquire more text data and improve language model training
Key Insight
💡 ASR pipelines can be effective in acquiring text data for low-resource languages, with a significant reduction in WER
Share This
🗣️ Use ASR pipelines to acquire text data for low-resource languages! 📊 Evaluate their effectiveness with metrics like WER
Key Takeaways
Learn how to use ASR pipelines to acquire text data for low-resource languages like Fongbe and Hausa, and evaluate their effectiveness
Full Article
Title: From Speech to Text Corpora: Evaluating ASR-Based Data Acquisition for Low-Resource Fongbe and Hausa
Abstract:
arXiv:2606.22274v1 Announce Type: cross Abstract: Low-resource African languages lack text corpora needed for language model training. We investigate whether ASR pipelines can extend text resources for two typologically distinct West African languages: Fongbe (tonal, diacritic-rich) and Hausa (non-tonal). We fine-tune MMS-300M on a curated 12.3-hour Fongbe dataset, achieving 9.48% WER on the ALFFA benchmark - a 78% relative reduction from the prior 44.04% baseline - while preserving tonal diacri
Abstract:
arXiv:2606.22274v1 Announce Type: cross Abstract: Low-resource African languages lack text corpora needed for language model training. We investigate whether ASR pipelines can extend text resources for two typologically distinct West African languages: Fongbe (tonal, diacritic-rich) and Hausa (non-tonal). We fine-tune MMS-300M on a curated 12.3-hour Fongbe dataset, achieving 9.48% WER on the ALFFA benchmark - a 78% relative reduction from the prior 44.04% baseline - while preserving tonal diacri
DeepCamp AI