GenSEC-LLM/SLT-Task2-Post-ASR-Speaker-Tagging

Dataset Name: Dataset for ASR Speaker-Tagging Corrections (Speaker Diarization)

Description

• This dataset is pairs of erroneous ASR output and speaker tagging, which are generated from a ASR system and speaker diarization system. Each source erroneous transcription is paired with human-annotated transcription, which has correct transcription and speaker tagging.
• SEGment-wise Long-form Speech Transcription annotation (SegLST), the file format used in the CHiME challenges

Example) session_ge1nse2c.seglst.json

[
...
    {
        "session_id": "session_ge1nse2c",
        "words": "well that is the problem we have erroneous transcript and speaker tagging we want to correct it using large language models",
        "start_time": 181.88,
        "end_time": 193.3,
        "speaker": "speaker1"
    },
    {
        "session_id": "session_ge1nse2c",
        "words": "it seems like a really interesting problem I feel that we can start with very simple methods",
        "start_time": 194.48,
        "end_time": 205.03,
        "speaker": "speaker2"
    },
...
]

Structure

Data Split

The dataset is divided into training and test splits:

• Training Data: 222 entries
  • 2 to 4 speakers in each session
  • Approximately 10 ~ 40 mins of recordings
• Development Data: 13 entries
  • 2 speakers in each session
  • Approximately 10 mins of recordings
• Evaluation Data: 11 entries
  • 2 speakers in each session
  • Approximately 10 mins of recordings

Keys (items)

• session_id: "session_ge1nse2c",
• words: Transcription corresponding to the time stamp (start, end).
• start_time: Start time in second.
• end_time: End time in second.
• speaker: Speaker tagging in string "speaker<N>"

Source Datasets

err_source_text: This is the erroneous ASR-Diarization results to be fixed. Has dev, eval folders ref_annotated_text: This is the human annotated ground-truth for evaluation. Only dev split is included.

• Training Sources:

  • dev: 222 sessions

• Development Sources:

  • dev: 13 sessions

• Evaluation Sources:

  • eval: 11 Sessions

Access

The dataset can be accessed and downloaded through the HuggingFace Datasets library (i.e., This Repository).

Evaluation

This dataset can be evaluated by MeetEval Software

From PyPI

pip install meeteval

From source

git clone https://github.com/fgnt/meeteval
pip install -e ./meeteval

Evaluate the corrected segLST files:

python -m meeteval.wer cpwer -h err_source_text/dev/session_ge1nse2c.json -r ref_annotate_text/dev/session_ge1nse2c.json

Or after installation, you can use the following command alternatively.

meeteval-wer cpwer -h err_source_text/dev/session_ge1nse2c.json -r ref_annotate_text/dev/session_ge1nse2c.json

References

@inproceedings{park2024enhancing,
  title={Enhancing speaker diarization with large language models: A contextual beam search approach},
  author={Park, Tae Jin and Dhawan, Kunal and Koluguri, Nithin and Balam, Jagadeesh},
  booktitle={ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
  pages={10861--10865},
  year={2024},
  organization={IEEE}
}

@InProceedings{MeetEval23,
  title={MeetEval: A Toolkit for Computation of Word Error Rates for Meeting Transcription Systems},
  author={von Neumann, Thilo and Boeddeker, Christoph and Delcroix, Marc and Haeb-Umbach, Reinhold},
  booktitle={CHiME-2023 Workshop, Dublin, England},
  year={2023}
}