DONDO: Open w2v-BERT Speech-Recognition Base Models for African Languages

Automatic speech recognition (ASR) for African languages is often hindered by a lack of transcribed audio and the fragmentation of existing datasets. The DONDO (Democratizing Oral Neural Dialect Ontology) project addresses this by releasing a suite of 21 monolingual and 5 multilingual ASR base models covering 27 African language varieties. By leveraging the w2v-BERT 2.0 self-supervised encoder and training on license-clear religious texts, the researchers provide a foundation for developers to build speech-enabled applications for millions of speakers.

The authors employ a pragmatic, reproducible fine-tuning recipe. To overcome the performance degradation typically seen when training multilingual models on heterogeneous data, they use a two-step (and occasionally three-step) learning-rate-annealed schedule. First, the encoder undergoes coarse adaptation to the full multilingual mixture at a higher learning rate. Second, the model is annealed at a significantly lower learning rate, which effectively recovers and often surpasses monolingual performance benchmarks.

Additionally, the paper introduces a lightweight language-conditioning mechanism. By prepending a one-hot language identity as a sequence of prefix frames to the acoustic features, the researchers enable a single multilingual checkpoint to be steered to a specific target language at inference time without requiring architectural modifications.

DONDO represents a shift toward treating African languages as first-class subjects in speech technology. By releasing these models under the Apache-2.0 license, the project removes barriers to commercial and research-based deployment. These models serve as a shared, open test bed for the community to experiment with low-resource adaptation and efficient fine-tuning techniques, potentially benefiting a population of over 100 million first-language speakers.