Afrivoice: Multimodal Image and Speech Dataset for Six African Languages

A multimodal dataset from DigitalUmuganda, last updated in 2026, where each data point consists of a JPEG image, a corresponding audio WAV file describing the image, and often a transcription of the audio. The description lists over 500 audio hours per language for six languages: Shona, Lingala, Fulani, Malagasy, Wolof, and Somali, with over 100 transcribed hours each.

Use Cases

• Train automatic speech recognition models based on the transcribed audio files for African languages.
• Develop image-to-speech or image captioning systems based on the paired image and audio description data.
• Create or fine-tune multimodal AI models based on the alignment of visual content and spoken language.
• Conduct linguistic research on the listed African languages based on the audio and transcription data.

Strengths

• Includes over 500 hours of audio per language for six distinct African languages.
• Provides over 100 hours of transcribed audio per language, which can be used for supervised training.
• Each data point is multimodal, pairing an image with a descriptive audio file.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• The dataset page indicates a last updated date of 2026-05-25; freshness should be verified.

Provenance

Source

DigitalUmuganda on Hugging Face.

Freshness

Last updated 2026-05-25 13:38:21.

Geography

Likely covers regions where the listed languages (Shona, Lingala, Fulani, Malagasy, Wolof, Somali) are spoken.