This collection contains benchmark datasets created for speech recognition for African languages by the UCT NLP research group.
AI & ML interests
We develop language resources and language models for low-resource South African languages
Recent Activity
View all activity
OmniASR CTC models (300M, 1B, 3B) fine-tuned for isiZulu speech recognition, with training-data ablations.
-
uctnlp/omniASR-CTC-300m-v2-Zulu-Baseline
Automatic Speech Recognition • 0.3B • Updated • 30 -
uctnlp/omniASR-CTC-300M-v2-Zulu-Lwazi
Automatic Speech Recognition • 0.3B • Updated • 54 -
uctnlp/omniASR-CTC-300M-v2-Zulu-No-Lwazi
Automatic Speech Recognition • 0.3B • Updated • 12 -
uctnlp/omniASR-CTC-300M-v2-Zulu-All-v1
Automatic Speech Recognition • 0.3B • Updated • 7
Models adapted for isiXhosa, isiZulu, isiNdebele, and Siswati via multilingual continued pretraining.
This collection contains benchmark datasets created for speech recognition for African languages by the UCT NLP research group.
OmniASR CTC models (300M, 1B, 3B) fine-tuned for isiZulu speech recognition, with training-data ablations.
-
uctnlp/omniASR-CTC-300m-v2-Zulu-Baseline
Automatic Speech Recognition • 0.3B • Updated • 30 -
uctnlp/omniASR-CTC-300M-v2-Zulu-Lwazi
Automatic Speech Recognition • 0.3B • Updated • 54 -
uctnlp/omniASR-CTC-300M-v2-Zulu-No-Lwazi
Automatic Speech Recognition • 0.3B • Updated • 12 -
uctnlp/omniASR-CTC-300M-v2-Zulu-All-v1
Automatic Speech Recognition • 0.3B • Updated • 7
MzansiText, an open pretraining corpus, and MzansiLM, a 125M decoder-only model, for all eleven official South African languages.
Models adapted for isiXhosa, isiZulu, isiNdebele, and Siswati via multilingual continued pretraining.