Building speech recognition for Ghanaian languages starts with one thing: data. Today we’re introducing ghana-speech — a dataset with 1,411,467 audio segments totaling 2,247 hours across 42 Ghanaian languages.
This isn’t a small experimental dataset. This is the scale needed to train real, production-quality ASR systems.
The dataset covers 42 languages with varying amounts of audio. Here are the top languages by hours:
| Language | Hours |
|---|---|
| Asante Twi | 200h |
| Ewe | 161h |
| Hausa | 153h |
| Fante | 122h |
| Kasem | 98h |
| Lelemi | 94h |
| Bimoba | 91h |
| Gonja | 86h |
| Buli | 86h |
| Sisaala Tumulung | 82h |
Every language has at least 3 hours of audio, with most having 15+ hours.
The backbone is Bible audio recordings. Why Bible data? Because every clip has a verified text transcription, the recordings are clean professional audio, and Bible translations exist for almost every Ghanaian language.
We supplemented this with other sources to add speaking style diversity.
For ASR Development: Previous efforts had tens or hundreds of hours. With 2,247 hours, you can now train serious models that work in real-world conditions.
For Transfer Learning: Train on a well-resourced language like Twi, then fine-tune on a lower-resource language. Acoustic patterns transfer well across related languages.
For Everyone: This dataset is open source (CC-BY-NC-4.0) on Hugging Face. No paywalls, no restrictions.
from datasets import load_dataset
ds = load_dataset("ghananlpcommunity/ghana-speech", "Asante_Twi_twi", split="train")
print(ds[0])