wouk1805/medreport_audio_204

MedReport - Audio Dataset

Dataset Description

This dataset contains medical report audio files with their transcriptions, formatted according to HuggingFace Audio Dataset specifications. It's suitable for training speech-to-text models and instruction-following models in the medical domain.

Dataset Structure

This dataset follows the official HuggingFace Audio Dataset format:

dataset/
└── train/
    ├── audio/
    │   ├── 20240315143022.wav
    │   ├── 20240315143155.wav
    │   └── ...
    ├── metadata.csv
    └── metadata.jsonl

Columns

• file_name: Relative path to audio file (e.g., "audio/20240315143022.wav")
• text: Human transcription of the audio (primary transcription column)
• id: Audio file identifier (filename only, e.g., "20240315143022.wav")

Statistics

• Total audio files: 204
• Format: WAV
• Metadata formats: CSV and JSONL
• License: Apache License 2.0
• Created: 2025-08-05

Usage

Loading the dataset

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("wouk1805/medreport_audio_204")

# Access audio and transcription
for example in dataset['train']:
    audio = example['audio']
    transcription = example['text']
    print(f"Transcription: {transcription}")

For speech-to-text training

# Clean, simple dataset structure
audio_column = "audio"
text_column = "text"
id_column = "id"  # For tracking/identification

Files

• train/audio/: Directory containing all WAV audio files
• train/metadata.csv: Main metadata file (HuggingFace standard)
• backup_metadata.jsonl: Alternative JSONL format (for reference)

Citation

@dataset{medreport_audio_dataset,
  author = {Young-wouk KIM},
  title={MedReport - Audio Dataset},
  year={2025},
  url={https://huggingface.co/datasets/wouk1805/medreport_audio_204}
}