SOREVA Multilingual Speech Dataset for African Languages

SOREVA is a multilingual speech dataset designed for evaluating text-to-speech and speech representation models. It contains approximately 150 audio and transcription samples for each of 49 African languages and dialects. The dataset was created by OlameMend and last updated in December 2025.

Use Cases

• Benchmarking text-to-speech model performance on out-of-domain African languages using the provided audio samples and transcriptions.
• Evaluating speech representation learning models on the 49 distinct languages and dialects included.
• Training or fine-tuning speech models for low-resource language generalization using the paired audio-transcription data.

Strengths

• Covers 49 distinct African languages and dialects.
• Provides approximately 150 paired audio and transcription samples per language.

Limitations

• Small sample size per language (approx. 150).
• Specific geographic distribution and dialect coverage within languages is unknown.

Provenance

Source

Goethe Institut initiative.

Collection Method

Collection of audio samples and transcriptions.

Freshness

Last updated in December 2025.

Geography

African languages and dialects.