mGhana-ST, Best African Dataset at Deep Learning Indaba 2026: Speech Translation Data for Twi, Fante, Ewe, and Ga

mGhana-ST is an open speech translation corpus for Ghanaian languages, covering Akan in its Akuapem Twi, Asante Twi, and Fante varieties, alongside Ewe and Ga.

AdwumaTech built it from native speaker recordings in Ghana and published it on Hugging Face under a permissive licence. Each item pairs source-language audio with English text, varieties are held as separate splits, and recordings carry non-verbal event tags. In August 2026 it was named Best African Dataset at the Deep Learning Indaba in Lagos, Nigeria.

Open dataset release is standard practice across speech and language research, and the interesting questions about a corpus are technical. What architectures does it unlock. What can be measured with it that could not be measured before. What class of system becomes buildable that was previously out of reach. This piece answers those three.

Speech translation data is the constrained artifact

Speech data is usually discussed as a single quantity. It is a ladder, and each rung supports a different class of system.

Almost all available African language speech data sits on the third rung. It supports transcription and stops there. The fourth rung is where a system that takes Ghanaian speech and returns actionable English becomes trainable, and it is the rung that requires a translation pass over every utterance during collection.

That extra pass is what makes speech translation data scarce, and it is why mGhana-ST sits at the fourth rung by design.

What is inside mGhana-ST

Four construction properties carry more technical weight than corpus size.

Each item carries a translation. Source audio is paired with English text, which supports end-to-end speech translation training directly with no intermediate transcription step required at inference.

Akan is split by variety. Akuapem Twi, Asante Twi, and Fante are held apart. Collapsing Akan under one label produces degraded performance for whichever variety runs thinnest in the training mix, and an aggregate Akan test set will not surface it.

Recordings carry non-verbal event tags. Laughter, hesitation, overlapping speech, and background conditions are marked and retained. Production audio contains all of it.

Recording conditions match deployment conditions. Native speakers were recorded in Ghana across varied acoustic environments, with the accents, register, and code-switching patterns of the population the resulting systems serve.

The corpus is published on Hugging Face at adwumatech-ai/mghana-st under a permissive licence and has passed 15,000 cumulative downloads. AdwumaTech maintains it as a standing open release alongside Dagbani and Dagaare.

The architecture the data determines

The default approach to translating speech is a cascade. A recognition model transcribes the audio, then a translation model translates the transcript. It assembles from existing components, and in low-resource conditions it fails in a specific, compounding way.

Cascades propagate error. Whatever the recognition stage gets wrong becomes ground truth for the translation stage, which has no access to the acoustic evidence that would have resolved the ambiguity. A cascade's end-to-end accuracy is bounded by the product of its stages, so two components at 85 percent yield roughly 72 percent through the full path.

End-to-end speech translation removes the intermediate transcript and the error riding on it. It requires paired audio and target-language text, which is the fourth rung of the ladder. Without that data the architecture is unavailable, and a team is forced into the cascade regardless of whether the cascade suits the deployment.

This is the practical meaning of a data gap. It removes design options.

Tone, and why the metrics have to change

Akan, Ewe, and Ga are tonal languages. Pitch carries lexical and grammatical meaning, and pipelines developed on non-tonal languages systematically under-model that channel. The resulting errors are not distributed noise. They are changes in meaning, which means they pass fluency checks and fail correctness checks.

Evaluation practice compounds this. BLEU penalises the morphological variation that is normal in agglutinative and tonal languages, which makes it a weak instrument for measuring progress in exactly the languages that most need measurement. Character-level metrics such as chrF and chrF++ track human judgement more closely across these language families.

AdwumaTech evaluates African language translation on character-level metrics alongside human adequacy judgement by native speakers, reported per variety, and treats a BLEU-only claim as unmeasured.

What the capability gap measures

On an internal AdwumaTech evaluation set of Akan and Ewe task outputs, a general purpose frontier model returned a task error rate of 38.2 percent. At the review rates a qualified native-speaker reviewer sustains, that generates roughly 74 hours of human review per thousand outputs.

That figure describes a system property. A deployment operating at that error rate requires a standing human review function sized to its throughput, which places a hard ceiling on the volume the system can serve and makes unattended operation unavailable. The model is capable in the abstract. The deployment fails at the language boundary, and the boundary exists where training data was never collected.

Corpus hours move that number directly. There is no architectural substitute for labelled in-language data at the point where a system meets a speaker.

What this enables in sovereign deployment

Sovereign AI deployment is usually discussed in terms of infrastructure location and data residency. Those are necessary conditions. The sufficient condition is system capability in the languages the state actually serves, and that capability has technical prerequisites the corpus layer supplies.

In-language citizen interfaces become buildable. A citizen services system that answers in Twi, Fante, Ewe, or Ga requires in-language understanding at measured quality.

Per-language performance becomes measurable, which is an assurance requirement. Public sector AI systems have to produce disaggregated evaluation evidence, measured across the populations they serve.

Models can be trained and retrained in country. A state that holds the data can train, evaluate, and retrain its own models on its own infrastructure.

Smaller in-language models become viable at the edge. Task-specific models trained on targeted in-language data run at a fraction of the parameter count that general purpose multilingual models require.

Capability transfers across the region. Variety-level handling learned across Akan transfers to other dialect continua, and collection methodology ports to neighbouring language families.

Licence terms as a deployment constraint

Licence choice determines whether a corpus can enter a production system, which makes it a technical parameter.

mGhana-ST carries permissive terms. A research-only release closes the publication gap and leaves the deployment gap in place. AdwumaTech releases open language data under terms permitting commercial deployment so the corpus is available at the layer where systems get built.

The work this sits inside

mGhana-ST contributes to an established body of African language resources. Masakhane's community-built datasets established distributed African NLP research at scale. The Lacuna Fund financed African language dataset creation across the continent. BibleTTS and OpenSLR put substantial speech under commercially usable terms for several West African languages. University of Ghana researchers built open ASR corpora across Akan, Ewe, Dagbani, Dagaare, and Ikposo.

AdwumaTech works alongside that ecosystem through formal academic MOUs with the University of Ghana and Valley View University, under which computer science and data science students move through structured programmes into annotator, trainer, and engineering roles inside live engagements.

What AdwumaTech builds on it

A public corpus supplies the acoustic and general-domain foundation. Production systems fail past that boundary, and the failure points are predictable enough to design collection around.

Domain vocabulary and entity handling. General-domain speech data does not cover the terms a deployed system meets. In financial services, numerals, currency amounts, dates, account identifiers, and institution names carry the highest cost per error and the lowest coverage in general corpora.

Code-switching. Ghanaian speech alternates between local languages and English continuously, frequently inside a single clause. AdwumaTech collects and annotates code-switched speech as a first-class category.

Task and preference data. Recognition and translation are the input path. A deployed system also has to respond, which requires supervised task data and preference data in-language.

The collection pipeline itself. The reusable asset is the machinery: annotation specifications, speaker recruitment, acoustic condition sampling, inter-annotator agreement thresholds, and adjudication protocols.

The evaluation harness. AdwumaTech delivers per-language and per-variety held-out test sets alongside the training data, with character-level metrics and human adequacy judgement defined as the measurement protocol.

That work runs through the data operations practice and the African languages practice. The same capability runs inside national systems, where in-language quality is a measured requirement carried per language and per variety.

Coverage, and what remains

mGhana-ST covers Akan across three varieties, Ewe, and Ga. AdwumaTech's open releases extend to Dagbani and Dagaare. Ghana alone has substantially more languages in active use, and the continent has over two thousand, of which fewer than fifty have meaningful representation in production AI systems.

The gap is addressable through collection. Each additional variety brought to the fourth rung of the ladder makes a further class of system buildable for the population that speaks it, and the methodology for getting there is established and repeatable.

Related reading: Which AI Use Case Should You Build First covers how AdwumaTech prices the review cost of a model operating at the edge of its evaluated performance. AI Assurance in the Public Sector covers why disaggregated evaluation is an evidence requirement in national systems.

Bring the problem. Keep the system. Talk to us.

References

• mGhana-ST dataset, AdwumaTech AI, Hugging Face: huggingface.co/datasets/adwumatech-ai/mghana-st
• Deep Learning Indaba 2026, Pan-Atlantic University, Lagos, 2 to 7 August 2026.
• Masakhane: MasakhaNER, MasakhaPOS, and MAFAND-MT African language datasets.
• Lacuna Fund, African language dataset portfolio.
• BibleTTS and OpenSLR West African speech releases.
• University of Ghana, open ASR audio datasets for Akan, Ewe, Dagbani, Dagaare, and Ikposo.

About AdwumaTech

AdwumaTech AI engineers applied AI systems for enterprises and sovereign digital infrastructure for governments, building the full chain from data acquisition through post-training to the deployed system. AdwumaTech collects, annotates, and open-sources speech and text datasets in Ghanaian languages including Twi, Fante, Ga, Ewe, Dagbani, and Dagaare. The enterprise practice covers AI consulting, productized AI, data operations, and NOKORE AI identity integrity. The government practice covers digital identity, payments infrastructure, citizen services, and AI assurance. AdwumaTech is ISO 27001 certified and aligned to ISO/IEC 42001, operates from Accra, Ghana, and holds formal academic MOUs with the University of Ghana and Valley View University.