README.md

metadata

language:
  - en
license: mit
task_categories:
  - automatic-speech-recognition
tags:
  - medical
  - audio
  - speech-recognition
pretty_name: 'YouTube Medical Lessons ASR Benchmark '
size_categories:
  - n<1K

Dataset Card for Medical Lessons Speech Corpus

Table of Contents

• Dataset Description
  • Dataset Summary
  • Supported Tasks
  • Languages
• Dataset Structure
• Dataset Creation
  • Curation Rationale
  • Annotations
• Considerations for Using the Data
• Additional Information
  • Citation Information

Dataset Description

Dataset Summary

The Medical Lessons Speech Corpus is a specialized audio dataset designed to benchmark Automatic Speech Recognition (ASR) systems on highly technical, domain-specific language. It contains a collection of medical lesson audio files accompanied by their corresponding transcriptions.

A defining characteristic of this dataset is its high density of complex medical terminology, making it an excellent stress test for evaluating the vocabulary coverage and phonetic accuracy of speech-to-text models. To ensure the highest standard of evaluation reliability, all transcripts have been strictly manually verified by the author, providing a highly accurate ground truth.

This dataset was originally created for the paper WhisperX with Context-Aware Transcription to evaluate state-of-the-art ASR models on challenging medical domain speech.

Supported Tasks

• automatic-speech-recognition: The primary task is to accurately transcribe the spoken medical lessons into text. The dataset serves as a robust benchmark for evaluating domain-specific ASR performance.

Languages

The audio and transcriptions are in English (en).

Dataset Structure

Data Instances

A typical instance consists of an audio file and its corresponding ground truth transcript file in JSON format. The JSON transcript contains an array of speech segments with text and timing information.

Audio Format: Standard audio file (e.g., WAV or MP3). Transcript Format (.json):

[
  {
    "text": " Hey, everyone. Here's another lesson in the medical terminology the basic series...",
    "start": 0.031,
    "end": 21.04,
  },
  {
    "text": " The next one is acro, this means extremity...",
    "start": 22.137,
    "end": 51.753,
  }
]

Data Fields

• text: The manually verified text transcript for the speech segment.
• start: The start time of the speech segment in seconds.
• end: The end time of the speech segment in seconds.

Dataset Creation

Annotation process

The ground truth transcripts were initially generated by openai whisper-1 model and then manually verified and corrected by the author. This rigorous manual review process ensures that all complex medical terms, drug names, and anatomical references are spelled accurately, providing a gold-standard reference for benchmarking.

Who are the annotators?

The manual verification was performed by the author of the paper XYZ, ensuring domain expertise in reviewing the medical terminology.

Considerations for Using the Data

Social Impact of Dataset

By providing a high-quality benchmark for medical ASR, this dataset encourages the development of more accurate speech-to-text systems for healthcare. Improvements in this domain can lead to better clinical documentation tools, enhanced accessibility for medical education, and reduced administrative burden for healthcare professionals.

Additional Information

Citation Information

If you use this dataset in your research, please cite the following paper:

@article{Author202X,
  title={WhisperX with Context-Aware Transcription},
  author={Carlos Bain},
  year={2026}
}