UGAkan-ImpairedSpeechData (Research Working Copy)

Personal working copy for ASR research. Original dataset: UGAkan-ImpairedSpeechData (University of Ghana).

Structure on this repo

Metadata Columns

Column	Description
filename	Audio filename
aetiology	Raw aetiology label
aetiology_key	Normalised key
gender	Speaker gender
speaker_id	Unique speaker ID
environment	Recording environment
transcription	Akan transcription
duration	Audio duration in seconds
severity	Severity label
audio_path	Original Drive path
split_unified	train / val / test_clean / test_hard
split_aetiology	Per-aetiology split for cross-generalization

Split Strategy

• Split by speaker ID — no leakage
• Unified: 70% train / 10% val / 10% test_clean / 10% test_hard
• Per-aetiology: separate splits for cross-generalization experiments
• Stroke = zero-shot test condition (male only)