DONDO: Open w2v-BERT Speech-Recognition Base Models for African Languages
DONDO releases Apache-2.0 ASR base models covering 27 African language varieties.
The paper describes 21 monolingual and five multilingual w2v-BERT 2.0 models across languages in Ghana, Sierra Leone, Nigeria, Senegal, Kenya, and Zimbabwe. Training uses mostly read speech from religious texts, chosen for license-clear and consistent transcriptions. The multilingual models use annealed fine-tuning and language-conditioning prefix frames, reaching average WERs of 10-13% across the five families. The models are posted under the Hugging Face KhayaAI organization for free fine-tuning, including commercial use. ArXiv · AI/CL/LG's note
The paper describes 21 monolingual and five multilingual w2v-BERT 2.0 models across languages in Ghana, Sierra Leone, Nigeria, Senegal, Kenya, and Zimbabwe. Training uses mostly read speech from religious texts, chosen for license-clear and consistent transcriptions. The multilingual models use annealed fine-tuning and language-conditioning prefix frames, reaching average WERs of 10-13% across the five families. The models are posted under the Hugging Face KhayaAI organization for free fine-tuning, including commercial use. ArXiv · AI/CL/LG's note
score 5