OCW Contextual ASR

Dataset Summary

OCW Contextual ASR is a lecture-based dataset for contextual automatic speech recognition research in the education domain. It is constructed from two MIT OpenCourseWare courses and is intended to support the evaluation of ASR with external contextual information such as lecture slides and named entities. The dataset is introduced in the following paper: Improving Named Entity Transcription with Contextual LLM-based Revision（https://arxiv.org/abs/2506.10779）

This release contains utterance-level metadata (transcription), named entity annotations, Whisper large predictions, and lecture materials organized into development and test splits.

Splits and Source Data

• dev: Brain Structure and Its Origins (https://ocw.mit.edu/courses/9-14-brain-structure-and-its-origins-spring-2014/)
  • 35 lecture audio recordings
• test: Animal Behavior (https://ocw.mit.edu/courses/9-20-animal-behavior-fall-2013/)
  • 25 lecture audio recordings
  • totaling 17 hours of audio
  • transcription contains approximately 121K words
  • the context consists of 46K words
  • ground truth transcriptions contain 1.2K named entity words

Included Files

This dataset currently includes:

• dev_utterance_metadata.csv
• dev_utterance_entities.csv
• dev_whisper_large_predictions.csv
• test_utterance_metadata.csv
• test_utterance_entities.csv
• test_whisper_large_predictions.csv
• dev.zip
• test.zip

File Description

1. *_utterance_metadata.csv

Utterance-level metadata with the following columns:

• audio_start_sec: utterance start time in seconds
• audio_filepath: relative path to the utterance audio file
• duration: utterance duration in seconds
• text: reference transcription text

2. *_utterance_entities.csv

Utterance-level named entity annotations with the following columns:

• audio_filepath: relative path to the utterance audio file
• whisper_large_entities: named entities extracted from Whisper large predictions
• transcript_entities: named entities extracted from the reference transcript

3. *_whisper_large_predictions.csv

Utterance-level Whisper large predictions with the following columns:

• audio_filepath: relative path to the utterance audio file
• whisper_large_pred_text: Whisper large predicted text
• whisper_large_word_probs: word-level confidence scores

4. *.zip

Compressed lecture-level materials. These include files such as:

• uttrances: .flac files
• lecture transcript text
• lecture transcript PDF
• lecture slides PDF

Licensing and Usage Notes

This dataset is derived from MIT OpenCourseWare materials. Users should also consult the original source pages for the corresponding course materials and license terms.

This release is intended for research use. Please check the copyright and license conditions of the original source materials before redistributing or reusing any raw content.

Citation

If you use this dataset, please cite our paper.

@article{trinh2026improvingnamedentitytranscription,
      title={Improving Named Entity Transcription with Contextual LLM-based Revision}, 
      author={Viet Anh Trinh and Xinlu He and Jacob Whitehill},
      year={2026},
      url={https://arxiv.org/abs/2506.10779}, 
      journal={Educational Data Mining},
}