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

Researchers have introduced DONDO, a new 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 spoken across Ghana, Sierra Leone, Nigeria, Senegal, Kenya, and Zimbabwe.

The models were primarily fine-tuned using read speech from religious texts. This approach was chosen due to the broad, license-clear, and orthographically consistent coverage these texts provide for languages that typically lack extensive transcribed audio datasets. The development involved a multi-step fine-tuning process, which first adapted a shared multilingual model at a higher learning rate before annealing it to achieve performance comparable to, and often surpassing, strong monolingual baselines.

A notable innovation is a lightweight language-conditioning mechanism. This allows a single multilingual model to be directed to a specific target language during inference by injecting a one-hot language identity as a prefix to the acoustic features. The annealed multilingual models achieved average word error rates (WER) between 10-13%, significantly closing the performance gap with monolingual models while consolidating multiple languages into a single checkpoint.

All DONDO models are released under the Apache-2.0 license on the Hugging Face KhayaAI organization. This permissive licensing allows for free use and fine-tuning, including for commercial applications. The languages covered by DONDO are estimated to be spoken by over one hundred million first-language speakers, with a substantially larger number including second-language users, highlighting the significant potential impact of these advancements for African communities.