Introducing African Speech ID
Our flagship language identification model supports 1,386 African languages and runs on CPU.
The Challenge
There are over 2,000 African languages, yet very few have been included in modern NLP research. Existing language identification models cover a fraction of these languages, leaving most African speech data unlabeled and unusable.
We set out to build a language ID model that covers as many African languages as possible, while remaining fast enough to run on consumer hardware.
Our Approach
African Speech ID uses a two-stage approach:
- Transcription: A fast version of Omnilingual ASR transcribes the audio into text characters.
- Classification: A lightweight classifier then identifies the language based on the transcription.
This approach is both accurate and efficient. The model runs at 3.5x realtime on a single CPU core, making it suitable for deployment on phones and laptops.
Results
The model covers 1,386 African languages including dialects. It requires just 5 seconds of audio and runs entirely on CPU with no GPU needed.