African Speech Recognition & Transcription Dataset

This is a synthetic dataset generated from high-quality expert-labelled seed data. All records are algorithmically derived — statistical distributions, inter-field correlations, and annotation characteristics faithfully replicate real-world patterns from the source data, while ensuring no real individual, organisation, or transaction can be identified or reconstructed.

The African Speech Recognition & Transcription Dataset spans 3,000+ hours of audio recorded across 12 African languages: Hausa, Yoruba, Igbo, Swahili, Zulu, Xhosa, Twi, Amharic, Wolof, Luganda, Kinyarwanda, and Mozambican Portuguese. Recordings were collected in controlled studio sessions, community centres, outdoor markets, and simulated call-centre environments — deliberately capturing the full acoustic range encountered in real-world deployment: background noise, varying microphone quality, speaker proximity variation, and multi-speaker overlap.

Each audio clip is paired with a human-verified orthographic transcript, a phonetic transliteration where applicable, speaker demographic metadata (age group, gender, dialect region), and acoustic environment tags. ASR confidence scores from a baseline wav2vec 2.0 model are included per clip, enabling curriculum learning approaches that sequence training from high-confidence to difficult utterances. Clip durations range from 2 to 30 seconds; the median is 8 seconds.

The dataset is partitioned into train / validation / test splits stratified by language, speaker identity (no speaker appears in both train and test), and acoustic environment. A separate out-of-domain evaluation set comprising radio broadcast excerpts and phone-call audio is provided for robustness testing. All audio is stored as 16 kHz mono WAV; transcripts are distributed as JSON sidecar files and as HuggingFace Datasets-compatible JSONL.