ALFFA Swahili Broadcast News Speech Corpus (SLR25)
~11.8 hours of RFI Swahili broadcast news audio, 10,180 train and 1,991 test utterances with manual transcripts at 16 kHz for ASR.
African research in artificial intelligence, machine learning, deep learning, generative AI, responsible AI, computer vision, natural language processing, data science, intelligent systems and AI applications for African development.
Repository handle: 123456789/18~11.8 hours of RFI Swahili broadcast news audio, 10,180 train and 1,991 test utterances with manual transcripts at 16 kHz for ASR.
31,000+ Swahili news articles in six topic classes (local, international, finance, health, sports, entertainment) with a fixed 75/25 train/test split.
Translations of a subset of MMLU into 15 African languages to benchmark multitask knowledge and reasoning.
A 0.4B-parameter transformer language model trained on African languages by Lelapa AI, released with associated research (arXiv:2408.17024).
~803k aligned four-way parallel sentences across English, Nuer, Dinka and Swahili for multilingual machine translation and cross-lingual transfer.
First large publicly available high-quality NER dataset for ten African languages, with PER/ORG/LOC/DATE annotations and train/dev/test splits.
Evaluation set of GSM8K translations into 16 African languages with 250 test problems per language for multilingual mathematical reasoning.
110,000+ annotated tweets for sentiment analysis across 14 under-represented African languages.
Largest publicly available topic-classification dataset based on Flores-200, covering 205 languages and dialects with a compact seven-class label set.
Speech version of FLoRes with ~2,009 n-way parallel sentences per language read by native speakers, for ASR, ST and language ID evaluation.
Largest MT benchmark for African languages in the news domain, with train/dev/test English-or-French to African-language translation pairs.
Practical content from the inaugural Deep Learning Indaba.