Dagbani Community Builds Voice AI Dataset in Tamale

The Dagbani Wikimedians User Group convened language volunteers in Tamale on 31 Jan - 1 Feb 2026 for a focused, community-centered workshop to produce annotated voice data using Mozilla Common Voice. Participants completed hands-on sessions covering sentence annotation, sentence validation, and quality-review workflows, and they contributed validated Dagbani sentence annotations intended for open voice datasets. The event explicitly targeted culturally appropriate phrasing, natural speech patterns, and consistency to improve downstream machine learning outcomes. The project also included Khmer in piloting and seeks to feed a larger dataset pool for training speech models and enabling an offline-capable, mobile-first application.

The workshop taught the operational steps practitioners must get right for small-language voice datasets. Key technical elements covered were Mozilla Common Voice annotation and validation pipelines, establishing quality-control criteria, and collaborative review practices. Participants practiced:

• •sentence-level annotation and validation workflows
• •consistency checks for linguistic and cultural accuracy
• •collaborative review rounds to surface natural phrasing and dialectal variation

The emphasis on validated, community-curated sentences reduces label noise and increases ecological validity for acoustic modeling and language modeling. For practitioners, the takeaways are that community annotation yields better lexical and prosodic coverage, but requires formalized guidelines, metadata capture (speaker demographics, recording conditions), and ongoing review to be ML-ready.

This workshop fits a broader movement toward community-driven datasets for underrepresented languages. Sparse, low-quality corpora are the main barrier to accurate ASR and TTS for African languages; community annotation events are a cost-effective way to bootstrap usable datasets while building local capacity. The combination of hands-on training plus contributions to an open platform like Mozilla Common Voice accelerates data availability for researchers and startups focused on low-resource speech models. The workshop also targets an operational constraint many practitioners face in emerging markets: offline deployment. By orienting the dataset and application design toward mobile-first, offline-capable experiences, the project aligns data collection with the real-world constraints of end users.

Practical implications for ML engineers: Expect data to initially be sentence-level, human-validated clips with richer cultural and dialectal coverage than scrape-based corpora. Key engineering questions to prepare for are dataset size thresholds for viable ASR baselines, strategies for speaker balance, augmentation plans to simulate noisy mobile recordings, and benchmark selection for evaluation. Projects aiming to build models from this data should codify annotation rules, define test splits that preserve speaker separation, and capture metadata necessary for downstream fairness and robustness analysis.

Sustainability and governance: The workshop model demonstrates how local communities can both collect and steward voice data, which improves dataset authenticity and long-term maintenance. Practitioners should note the need for continuous annotation, reviewer incentives, and a governance model that clarifies data licensing, contribution attribution, and privacy safeguards for voice data.

Ongoing annotation drives, a public dataset release on Mozilla Common Voice, early ASR/TTS baselines trained on the contributed Dagbani corpus, and the first prototypes of the offline mobile app. Continued funding, scalable contributor workflows, and tooling for quality control will determine whether the dataset moves from a pilot to a production-ready resource.