zionia/isindebele-asr

isiNdebele Speech Recognition Dataset

Dataset Description

This dataset contains speech recordings and transcriptions for isiNdebele, one of South Africa's official languages. The dataset has been optimized for use with OpenAI's Whisper ASR models.

Dataset Statistics

• Number of samples: 742
• Language: isiNdebele (Nbl)
• Audio format: WAV, 16kHz, mono, 16-bit
• Maximum duration: 30 seconds (truncated for Whisper compatibility)
• Transcription format: Cleaned text (lowercase, no punctuation)

Audio Processing

All audio files have been processed with the following optimizations:

• Resampled to 16kHz (Whisper's native sample rate)
• Converted to mono
• Normalized volume to -16dBFS
• Truncated to maximum 30 seconds
• 16-bit PCM encoding

Transcription Cleaning

Transcriptions have been cleaned:

• Converted to lowercase
• Removed all punctuation
• Removed POS markers (e.g., [n], [v])
• Normalized whitespace

Dataset Structure

Each sample contains:

• audio: Audio file path and array
• transcription: Cleaned transcription text
• file_id: Unique identifier for the recording
• subfolder: Original subfolder location

Usage

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("zionia/isindebele-lwazi-asr")

# Access samples
print(dataset['train'][0])

Citation

If you use this dataset, please cite the original Lwazi corpus:

@inproceedings{lwazi2011,
  title={The Lwazi corpus: an African speech resource},
  author={Barnard, E. and Davel, M. H. and Van Heerden, C.},
  booktitle={Proceedings of the 22nd Annual Symposium of the Pattern Recognition Association of South Africa},
  year={2011}
}

License

This dataset is released under the MIT License.

Language

isiNdebele is a language spoken primarily in South Africa.