Lwazi English Telephone Speech Corpus for ASR

South African English audio recordings and transcriptions for developing Lwazi speech recognition systems. The corpus contains telephone-quality audio files at 8 KHz, 16-bit, mono channel, with corresponding orthographic transcriptions in Unicode text format. Researcher Jaco Badenhorst created this dataset, which was last updated in April 2026.

Use Cases

• Train ASR models to map 8 KHz, 16-bit, mono-channel audio features to corresponding orthographic transcriptions.
• Benchmark speech recognition accuracy on telephone-quality audio using the provided utterance-level text transcriptions.
• Analyze phonetic or acoustic characteristics of South African English speech from the recorded utterances.
• Develop or fine-tune acoustic models for low-bandwidth, single-channel audio input common in telephony applications.

Strengths

• Audio recordings are standardized at 8 KHz, 16-bit, and 1 channel (telephone quality).
• Transcriptions are provided per utterance in Unicode text format, ensuring direct alignment for model training.

Limitations

• The dataset scale is limited, indicated by a file size of 0.0 B, suggesting a small sample for model training.
• No information is provided on the number of speakers, utterances, or total hours of audio, limiting reproducibility assessments.
• Geographic coverage is implied to be South Africa, which may limit generalizability to other English dialects.

Provenance

Source

Jaco Badenhorst via the Lwazi project.

Collection Method

Audio recordings and manual orthographic transcriptions collected for speech recognition system development.

Freshness

Last updated in April 2026.

Geography

South Africa (inferred from project name 'Lwazi').