Twi (Akan) Speech Dataset
Natural, code-switched Akan Twi speech for Ghanaian voice and ASR systems.
Ghana · ISO 639: ak
Listen to a sample
What you get with the Twi dataset
Part of our speaker-verified corpus across Luganda, Ewe, Fulfulde, Hausa, Wolof, Twi & Yoruba fully annotated and growing.
Built for teams training models that speak Twi
Twi ASR & TTS
Train speech recognition and synthesis on verified native Twi speech instead of scraped audio.
Multilingual & multimodal models
Add natural Twi as spoken in Ghana, code-switching included, to your pretraining and fine-tuning mix.
Evaluation sets
Measure how your models handle Twi on native-speaker validated test data.
How the data is made
Natural speech, not just read aloud
Read, spontaneous and conversational speech: we capture the language as it's actually spoken, code-switching included.
AI quality filtering at capture
Noisy audio is discarded automatically before it reaches human reviewers.
≥95% inter-annotator agreement
Every transcript is corrected by a native speaker, and a sample of every batch is blind double-annotated to measure agreement.
Fair-Trade sourcing
Native speakers are paid fairly per completed mission, with a provenance trail your compliance team can audit.
On Wolof sentiment, data validated through our pipeline reaches 90.0% F1, twice the best frontier LLM we tested.
See the benchmarkTwi dataset FAQ
Is the Twi speech dataset licensed for commercial use?
Yes. Every Twi dataset ships with a commercial license and an SLA: guaranteed delivery windows, defined annotation accuracy, free re-delivery on any QA failure, and a provenance trail for IP, data-protection and AI Act reviews.
What format is the audio delivered in?
16 kHz WAV audio with timestamped transcripts in JSON, delivered via API or S3 straight into your training pipeline. We can adapt to your schema.
How many hours of Twi audio are available?
The Twi corpus is growing continuously. Volume, speaker mix and domains are scoped to your use case on a short call, and we can run a collection mission to top up anything specific.
Can you build a Twi evaluation set for our models?
Yes. Beyond training data, we build native-speaker validated Twi test sets and benchmark your models against frontier baselines, as in our Wolof benchmark.
Can I hear a sample before committing?
Yes. Request the dataset or book a call and we will share representative samples and a data card before any contract.
Other shipping datasets
- Luganda Speech Dataset
- Ewe Speech Dataset
- Fulfulde (Fula) Speech Dataset
- Hausa Speech Dataset
- Wolof Speech Dataset
- Yoruba Speech Dataset
- Browse the full catalogue
Need another African language? We run on-demand collection missions with our native-speaker network.