New Open-Source AI Models Boost Speech Recognition for African Languages

Researchers have introduced DONDO, a collection of open-source automatic speech recognition (ASR) base models specifically designed for African languages. Built on the w2v-BERT 2.0 self-supervised speech encoder, DONDO includes twenty-one monolingual and five multilingual models, covering twenty-seven language varieties across six African countries: Ghana, Sierra Leone, Nigeria, Senegal, Kenya, and Zimbabwe.

The models were primarily fine-tuned using read speech from religious texts, a strategic choice to overcome the common challenge of limited transcribed audio for many African languages. This approach ensures broad, license-clear, and orthographically consistent data. The fine-tuning process involved a two-step learning-rate annealing procedure, allowing the models to adapt effectively and, in some cases, outperform existing monolingual baselines.

A notable innovation is the lightweight language-conditioning mechanism. This feature enables a single multilingual checkpoint to be directed to a specific target language during inference by injecting a one-hot language identity. This significantly improves efficiency and broad applicability.

The annealed multilingual models achieved impressive average word error rates (WER) of 10-13%, effectively closing the performance gap with monolingual models while consolidating support for numerous languages into a single checkpoint. This development is crucial for advancing AI accessibility and utility across the continent.

All DONDO models are openly available on the Hugging Face KhayaAI organization under the Apache-2.0 license, allowing free use and further fine-tuning, including for commercial applications. These models are estimated to serve approximately one hundred million first-language speakers, with a much larger reach when second-language users are considered, paving the way for more inclusive AI-powered applications in Africa.