This dataset contains 1 hour of South African–accented English speech paired with human-annotated transcripts. It is provided as a public sample to demonstrate the dataset structure, audio quality, and transcription style used in the full commercial version.

Language

English (South African accent)

Dataset Structure

Audio files: High-quality .mp3 format (48 kHz).
Metadata: Accompanied by a metadata.csv (or similar CSV) containing:
- ID: Unique identifier for each entry.
- File name: The associated audio file path.
- Transcript: The corresponding text transcription.

Transcription Details

Pipeline: Audio was initially processed using an Automatic Speech Recognition (ASR) system.
Human Review: Transcripts were subsequently reviewed and annotated by human annotators to ensure accuracy.
Real-world conditions: Transcriptions reflect real-world noise and imperfections, making the dataset suitable for noisy-label and robustness-focused training.

Intended Use

This public sample is intended for:

Dataset inspection and quality evaluation.
Accent analysis and linguistic research.
ASR experimentation and benchmarking.
Commercial assessment for licensing the full dataset.

Limitations

Sample Size: This is a small subset (1 hour) and is not intended for full-scale model training.
Error Rate: Transcripts are not 100% error-free. The estimated transcription error rate for the full dataset is approximately 15%.

Access & Licensing

This sample is publicly available for evaluation purposes. The full 300-hour dataset is owned by DATAI. For full access:

Contact: [email protected]

Downloads last month: 58