CDLI Non-standard speech datasets
Collection
Speech datasets of speakers with non-standard speech (different severity levels and types of speech impairment). • 6 items • Updated • 3
Datasets:
You need to acknowledge the terms and conditions to access the repository
We will review your request within 7 days.
Make sure that you have carefully read and understood the Terms_and_Conditions.md
Log in or Sign Up to review the conditions and access this dataset content.
This dataset provides 6.28 hours of Ga nonstandard speech recordings (12,160 samples) from 21 Ga speakers living with speech impairments. The participants represent a diversity of progressive, acquired and congenital aetiologies, including cerebral palsy, Parkinson's disease, multiple sclerosis, autism spectrum disorder, Down syndrome, stroke and stuttering.
This dataset includes a split into a training, test and development set. The splits were created avoiding any overlap on the speaker or phrase level. All speech recordings of this datasets have been transcribed by certified speech-language pathologists. Two passes of transcription have been made to ensure correctness. For this dataset, all recordings with low confidence transcriptions or disagreement between transcribers have been removed.
Additionally, this dataset contains speaker metadata such as gender, age, speech severity, impairment type, and underlying condition as assessed by certified speech-language pathologists. However, the dataset is fully anonymized, all potentially identifying information has been removed.
Recordings were gathered in and around Accra, Ghana in 2025. Recordings are based on picture and text-based prompts, designed to capture natural, everyday speech patterns.
This dataset was designed and commissioned by the Global Disability Innovation Hub and collected and operationalised by the University of Ghana, in collaboration with University College London and Talking Tipps Africa, through the Centre for Digital Language Inclusion (CDLI), as part of a project to support the development of automatic speech recognition tools for impaired speech in Ghanaian languages.
Users of this dataset must acknowledge our Terms & Conditions. We have anonymized the datasets to protect the speakers whose speech samples are contained in this dataset and it is strictly forbidden to attempt re-identify the speakers.
- Downloads last month
- 26