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

2026-07-25 · Source: cs.CL updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Speech Technology · Depth: Expert, long

Summary

DONDO is a new family of open, permissively licensed 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. These models are fine-tuned primarily using read speech from religious texts, chosen for their broad, license-clear, and orthographically consistent coverage in low-resource settings. A key innovation is a two-step (or three-step for one family) learning-rate-annealed fine-tuning procedure that adapts a shared multilingual model and then refines it to achieve average word error rates (WER) of 10–13%, often surpassing monolingual baselines. Additionally, a lightweight prefix-frame language-conditioning mechanism allows a single multilingual checkpoint to be steered to a target language at inference. All models are released under the Apache-2.0 license on Hugging Face KhayaAI, aiming to serve 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 consider fine-tuning these Apache-2.0 licensed w2v-BERT 2.0 base models on your specific in-domain data to achieve high-quality transcription. The included two-step annealing recipe and prefix-frame language conditioning simplify adapting multilingual models, potentially serving millions of speakers. This approach reduces development barriers and promotes localized speech technology.

Key insights

Open w2v-BERT 2.0 ASR base models for 27 African languages achieve competitive WERs via annealed fine-tuning and prefix-frame conditioning.

Principles

• Religious texts offer license-clear, consistent data for low-resource ASR.
• Annealed fine-tuning recovers and surpasses monolingual ASR baselines.
• Lightweight prefix-frame conditioning enables multilingual model steering.

Method

A two-step (optionally three-step) learning-rate-annealed fine-tuning procedure adapts a multilingual w2v-BERT 2.0 encoder, then refines it at a lower learning rate. Language conditioning uses prefix frames.

In practice

• Fine-tune DONDO models on in-domain data for specific use cases.
• Use prefix-frame conditioning to select target language at inference.
• Deploy multilingual models for regional language coverage.

Topics

• Automatic Speech Recognition
• African Languages
• w2v-BERT 2.0
• Low-Resource NLP
• Multilingual Models
• Self-supervised Learning
• Apache-2.0 License

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 ASR models for African languages achieve strong performance via annealed fine-tuning and language conditioning.
• Voices from the Margins: Modeling Linguistic Diversity in Spontaneous Speech for Low-Resource Languages — ASR for low-resource languages benefits from multilingual models and n-gram decoding, aiding revitalization and accessibility.
• AfriSUD: A Dependency Treebank Collection for Evaluating Models on African Languages — Models exhibit a significant "syntax gap" in processing African languages, struggling with their structural diversity.
• Expanding the Lexicon of Ge'ez Based African Languages: A Comparative Study of Amharic and Tigrinya — Vocabulary extension and two-stage training significantly boost multilingual PLM performance on low-resource, non-Latin-script languages.
• Optimization of Voice Translation Systems for Indigenous Languages: Retraining the NLLB-200 Model for the Quechua–Spanish Pair — Fine-tuning NLLB-200 with curated data and optimized training improves low-resource indigenous language translation on consumer hardware.

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 cs.CL updates on arXiv.org.