Join Us
Ghana Speech ID
  • Project : Ghana Speech ID
  • Company : Ghana NLP
  • Status : Active
  • Date : 21 August 2026

Interested to see how it works? Try Ghana Speech ID here:

Try the demo

What is Ghana Speech ID?

Ghana Speech ID tells you which language is being spoken, across 41 Ghanaian and West African languages. It is the piece that has to run before almost anything else: you cannot pick the right translation model, the right speech recogniser or the right voice until you know what language you are listening to.

How it works

The model does not read audio directly. Speech first goes through Omnilingual ASR, which turns it into a transcript; Ghana Speech ID then identifies the language of that text.

audio ──[sherpa-onnx + omniASR CTC]──▶ transcript ──[Ghana Speech ID]──▶ language

Splitting the work this way keeps the language identification itself tiny. The classifier is a vocabulary lookup and one sparse gather, not a neural network doing heavy arithmetic over audio.

Small enough to run anywhere

The model is 8.2 MB and runs on CPU only, in about 0.064 ms per classification with a 36 MB memory footprint. That is roughly four orders of magnitude cheaper than the speech recognition running in front of it — on any device that can already run the ASR, identifying the language is effectively free.

The inference core is C++ with a C API, so there is no Python on the device. There are JNI and Kotlin bindings for Android, and a Swift wrapper for iOS that imports the C API with no bridging header. The ONNX graph deliberately uses opset-13 core operators only, so it runs in mobile onnxruntime builds.

Give it enough audio

Accuracy depends more on how much speech you feed it than on anything else you control:

Audio length Accuracy
3 seconds 0.51
5 seconds 0.66
7 seconds 0.76
10 seconds 0.78

Five seconds is the practical minimum, ten gives the best result. Past ten seconds the curve flattens. It is also worth checking that most of the audio is actually speech before transcribing — a recording that is half silence carries half the evidence its duration suggests.

What it will and will not tell you

The model is closed-set: it always names one of its 41 languages, including for speech in a language it has never heard. It returns no answer at all when nothing in the transcript gave it a basis to decide, and that case should be reported as unknown rather than resolved to whichever language scored least badly.

The model card documents how the model behaves, what it scores, and where it fails. It is worth reading before relying on the confidence scores.

Using it

pip install ghana-speech-id

ghana-speech-id "obiara na enyi nyɛden dɛ ɔbɔbɔ no nkenyan"

The training and evaluation scripts in the repository reproduce the model end to end, and every release is checked for parity across scikit-learn, the Python package and the C++ runtime.

AISCA
Stanford University
UNICEF
Google
Microsoft
Harvard University