01Sources, rights & context
Data provenance
We consider where a resource originates, who contributed to it, the circumstances of its collection, and the permissions that govern its use.
Data & Research Governance
How we approach data provenance, research integrity, reproducibility, and responsible use of African speech and language resources.
Explore our approach01Our approach
Research quality begins with understanding the data.
The origins, composition, preparation, and permitted uses of a dataset shape what researchers can learn from it. These considerations are part of our work in African speech and language technology.
02Evidence, methods & limitations
Research integrity
We distinguish measured results from assumptions, document relevant methods, and account for limitations when interpreting research findings.
03Records, experiments & evaluation
Reproducibility
We value research records that explain how data was prepared, how experiments were conducted, and how evaluation results were obtained.
04Languages, speakers & context
Language representation
We consider the linguistic and social context of African language resources, including dialect variation, recording conditions, and gaps in coverage.
02The research lifecycle
Data governance is part of the research, not an afterthought.
Data decisions influence every stage of model development. Our approach considers a resource's origins, the transformations it undergoes, the evidence it supports, and the conditions under which it can be shared.
Source
We examine a resource's origin, collection context, intended use, and applicable permissions.
Prepare & annotate
We account for processing, filtering, transcription, labelling, and other transformations that shape the resulting data.
Train & evaluate
We consider whether the data and evaluation methods fit the research question and support the conclusions being drawn.
Assess limitations
We consider coverage gaps, uncertainty, data quality, and the conditions under which the results were obtained.
Document & maintain
We value records of methods, resource changes, corrections, and findings that help make research understandable.
Share responsibly
We consider permissions, privacy, attribution, and foreseeable risks when determining whether and how a resource can be shared.
03Research documentation
Make the work understandable.
Research records help explain how a resource was created, how an experiment was conducted, and where the evidence supports or limits a conclusion.
Dataset records
Sources, collection context, intended purpose, composition, applicable terms, and known limitations.
Processing & annotation
Preparation methods, filtering, normalisation, transcription, annotation decisions, and material transformations.
Experiments & evaluation
Relevant configurations, evaluation design, metrics, test conditions, and the evidence behind reported results.
Attribution & corrections
Contributions, attribution requirements, reuse conditions, known errors, and changes that affect interpretation.
04Language data needs context
African language data is not interchangeable.
A dataset reflects the people, language varieties, collection conditions, and annotation decisions behind it. These factors shape how a model performs and how far research findings can reasonably be generalised.
What we consider
- 01Language varieties and dialects
- 02Speaker representation
- 03Recording conditions and quality
- 04Annotation methods and consistency
- 05Privacy, consent, and permissions
- 06Limits on generalising results
05Data access & sharing
Openness depends on the resource.
We consider applicable permissions, privacy, attribution, research value, and foreseeable risks when deciding whether and how a resource can be shared.
Open access
Suitable for public sharing
Resources whose permissions and applicable conditions allow public release, with appropriate documentation and attribution.
Controlled access
Access with conditions
Resources whose legitimate use depends on additional permissions, access restrictions, or conditions on reuse.
Restricted release
Not suitable for public release
Resources that cannot responsibly be shared when permissions are absent or unclear, or when relevant risks remain unresolved.
06Research integrity
Claims are only as strong as the evidence behind them.
We distinguish demonstrated findings from hypotheses, recognise the contributions behind research resources, and account for relevant limitations when interpreting results.
Benchmark scores and successful demonstrations provide evidence under particular conditions. They do not, on their own, establish performance across every African language, dialect, speaker group, or real-world setting.
Questions about research or data?
For questions about a research resource, its provenance, permissions, or appropriate use, contact Sauti Labs through the contact channel published on our website. Include the relevant resource and a description of your question without sending unnecessary personal or sensitive information.