Turkana Speech Dataset: 529 Audio Clips from Bible Narratives

529 audio segments totaling 46 minutes provide speech data for Turkana, an Eastern Nilotic language with roughly 1 million speakers in Kenya. The dataset was created by Speedykom using Bible narratives from the Global Recordings Network, segmented via silence detection. Transcripts were auto-generated using the facebook/mms-1b-all model with a Teso adapter.

Use Cases

• Train automatic speech recognition models based on audio clips and corresponding transcripts.
• Fine-tune multilingual speech models for the Turkana language based on the provided audio-text pairs.
• Benchmark speech segmentation algorithms based on the silence detection method used to create the clips.
• Study prosody and phonetics in Turkana speech based on the 5.2-second average clip duration.

Strengths

• Contains 529 distinct audio segments, providing multiple data points for model training.
• Total duration of approximately 46 minutes offers a substantive audio corpus for a low-resource language.
• Audio is formatted as 16 kHz mono WAV files with UTF-8 transcripts, suggesting consistent preprocessing.

Limitations

• Transcriptions are auto-generated, which may introduce errors not present in human-verified data.
• Source material is limited to religious narratives from a single organization, which may not represent all speech domains.
• Column-level documentation and sample data are unavailable, making detailed quality assessment difficult prior to download.

Provenance

Source

Global Recordings Network (GRN) Bible narratives.

Collection Method

Audio segmented via silence detection; transcripts auto-generated via facebook/mms-1b-all (Teso adapter).

Freshness

Last updated 2026-04-15 15:05:08.

Geography

Kenya (language region).