Building Community-Centred NLP Resources for Puno Quechua

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Advanced, quick

Summary

New ASR resources have been developed for Puno Quechua (ISO 639-3: qxp), marking the first dedicated effort for this under-resourced language. This initiative produced the largest speech corpus for any single Quechua variety, comprising 66 hours of recordings, including 36 hours of manually transcribed and validated data. The data collection employed a participatory design campaign, ensuring community involvement. Additionally, the project established the first systematic ASR benchmark for Puno Quechua, evaluating state-of-the-art models such as Whisper-base, wav2vec2-base, and XLS-R-300M. These models were fine-tuned both with and without continued pre-training (CPT). All developed datasets and fine-tuned models are openly released to support further research and preservation efforts.

Key takeaway

For NLP Engineers or AI Scientists developing resources for other under-resourced languages, you should prioritize community-centered participatory design for data collection to ensure relevance and quality. Consider fine-tuning established models like Whisper-base or wav2vec2-base, evaluating the impact of continued pre-training. Crucially, openly releasing your datasets and fine-tuned models will significantly contribute to broader language preservation efforts and accelerate future research.

Key insights

Building NLP resources for under-resourced languages benefits significantly from community-centered participatory design and open data release.

Principles

• Participatory design is key for under-resourced language NLP.
• Openly releasing datasets and models aids preservation.
• Systematic benchmarking validates ASR model performance.

Method

Speech data was collected via a participatory design campaign, manually transcribed, and validated. ASR models like Whisper-base, wav2vec2-base, and XLS-R-300M were fine-tuned with and without continued pre-training, then benchmarked.

In practice

• Implement participatory design for data collection.
• Fine-tune Whisper-base or wav2vec2-base for ASR.
• Evaluate continued pre-training for performance gains.

Topics

• Puno Quechua
• Automatic Speech Recognition
• Low-Resource Languages
• Participatory Design
• Speech Corpora
• Whisper Model
• wav2vec2

Best for: Research Scientist, AI Scientist, NLP Engineer

Related on AIssential

• Giving Voice to the Constitution: Low-Resource Text-to-Speech for Quechua and Spanish Using a Bilingual Legal Corpus — Cross-lingual transfer in TTS effectively mitigates data scarcity for low-resource languages like Quechua.
• A Community Roadmap for Including Kapampángan in Modern AI Research — Community-driven data creation is essential for including underrepresented languages in AI research and applications.
• BaltiVoice: A Speech Corpus and Fine-tuned Whisper ASR System for the Balti Language — Fine-tuning pre-trained ASR models with modest, newly created corpora can establish functional speech recognition for low-resource languages.
• Paza: Introducing automatic speech recognition benchmarks and models for low resource languages — Human-centered design and evaluation are crucial for effective speech models in low-resource, real-world environments.
• Data Synthesis and Parameter-Efficient Fine-Tuning for Low-Resource NMT: A Case Study on Q'eqchi' Mayan — Synthetic data effectively bootstraps NMT structural learning for low-resource languages, but requires authentic data for semantic depth.
• Māori Text-to-Speech Model Spurns Big Tech’s Values — Indigenous communities can develop high-fidelity AI language models prioritizing data sovereignty and community ownership, offering an alternative to big tech.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.