Korean Speech Recognition Dataset With 51.6 Training Hours

The Zeroth-Korean dataset contains approximately 51.6 hours of training data and 1.2 hours of test data for Korean automatic speech recognition. It was created by kresnik and last updated in October 2024.

Use Cases

• Train an automatic speech recognition model on 51.6 hours of Korean audio training data.
• Benchmark ASR model performance using the 1.2 hours of provided test data.
• Fine-tune pre-trained speech models for the Korean language modality.

Strengths

• Contains 51.6 hours of dedicated training data for model development.
• Includes a separate 1.2-hour test set for evaluation.
• Publicly available under a permissive license for research and development.

Limitations

• The total dataset size of 52.8 hours is relatively small for modern speech model training.
• Specific audio characteristics, speaker demographics, and recording conditions are not detailed.
• The dataset's creation methodology and potential label noise are not described.

Provenance

Source

huggingface

Collection Method

Created for Korean ASR research, specific collection method unknown.

Freshness

Last updated on 2024-10-31.

Geography

South Korea (inferred from language focus)