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

2026-07-23 · Source: Computation and Language · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Advanced, quick

Summary

DONDO is a new family of open, permissively licensed automatic speech recognition (ASR) base models specifically designed for African languages, leveraging the w2v-BERT 2.0 self-supervised speech encoder. This release includes twenty-one monolingual and five multilingual models, covering twenty-seven language varieties across Ghana, Sierra Leone, Nigeria, Senegal, Kenya, and Zimbabwe. The models are fine-tuned primarily on read speech from religious texts, chosen for their broad, license-clear, and orthographically consistent coverage. A two-step (or three-step for one family) learning-rate-annealed fine-tuning procedure was employed, first adapting a shared multilingual model at a high learning rate, then annealing it to achieve average word error rates (WER) of 10-13% for multilingual models. A lightweight language-conditioning mechanism allows steering a single multilingual checkpoint to a target language at inference. All DONDO models are available on the Hugging Face KhayaAI organization under the Apache-2.0 license, supporting free commercial use for languages spoken by an estimated one hundred million first-language speakers.

Key takeaway

For Machine Learning Engineers developing ASR solutions for African languages, DONDO offers a robust, open-source foundation. You should leverage these Apache-2.0 licensed w2v-BERT models from Hugging Face KhayaAI to accelerate development for 27 language varieties. Consider applying the two-step annealed fine-tuning and lightweight language-conditioning mechanism to adapt models or build new multilingual systems, enabling efficient deployment for millions of speakers.

Key insights

Open w2v-BERT ASR models for African languages achieve strong performance via annealed fine-tuning and language conditioning.

Principles

• Religious texts offer license-clear data for low-resource languages.
• Annealed fine-tuning improves multilingual ASR performance.
• Language conditioning enables single-checkpoint multilingual inference.

Method

A two-step learning-rate-annealed fine-tuning procedure adapts a shared multilingual model at a high learning rate, then anneals it to recover or surpass monolingual baselines.

In practice

• Utilize DONDO models for ASR in 27 African language varieties.
• Implement language conditioning for multilingual ASR inference.
• Explore religious texts as data sources for low-resource NLP.

Topics

• Automatic Speech Recognition
• African Languages
• w2v-BERT
• Multilingual Models
• Self-supervised Learning
• Low-Resource NLP

Best for: AI Engineer, Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• DONDO: Open w2v-BERT Speech-Recognition Base Models for African Languages — Open w2v-BERT 2.0 ASR models for African languages achieve strong performance with a novel fine-tuning and language-conditioning approach.
• DONDO: Open w2v-BERT Speech-Recognition Base Models for African Languages — Open w2v-BERT 2.0 ASR base models for 27 African languages achieve competitive WERs via annealed fine-tuning and prefix-frame conditioning.
• Polyglot-Lion: Efficient Multilingual ASR for Singapore via Balanced Fine-Tuning of Qwen3-ASR — Balanced fine-tuning of moderate-scale ASR models yields cost-effective, high-performance multilingual speech recognition.
• Pretrained self-supervised speech models can recognize unseen consonants — Pretrained self-supervised speech models effectively generalize to and accurately recognize typologically rare phonemes, such as click consonants.
• From Monolingual to Multilingual: Evaluating Mamba for ASR in South African Languages — Mamba offers Conformer-level ASR accuracy with less compute, and multilingual training with control vectors improves robustness for low-resource languages.
• Voice Activation Detection for Transcription of Indigenous Languages — Hybrid neural VAD models like SpeechBrain perform best for Indigenous language fieldwork audio, despite challenges with non-speech segments.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computation and Language.