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

---
license: apache-2.0
---

# 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](#segment-wise-long-form-speech-transcription-annotation-seglst) (`SegLST`), the file format used in the [CHiME challenges](https://www.chimechallenge.org)

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](https://github.com/fgnt/meeteval)

### 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

```bib
@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}
}
```

```bib
@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}
}
```

原始信息汇总

数据集概述：ASR Speaker-Tagging Corrections (Speaker Diarization)

数据集描述

• 本数据集包含错误的自动语音识别（ASR）输出和说话人标记，这些数据由ASR系统和说话人分割系统生成。
• 每个错误的转录文本都与人工标注的转录文本配对，后者包含正确的转录和说话人标记。
• 使用SegLST文件格式，该格式用于CHiME挑战。

数据结构

数据分割

• 训练数据：222个条目，每个会话包含2至4个说话人，录音时长约10至40分钟。
• 开发数据：13个条目，每个会话包含2个说话人，录音时长约10分钟。
• 评估数据：11个条目，每个会话包含2个说话人，录音时长约10分钟。

关键字段

• session_id：会话标识符。
• words：与时间戳（开始，结束）对应的转录文本。
• start_time：开始时间，单位为秒。
• end_time：结束时间，单位为秒。
• speaker：说话人标记，格式为"speaker<N>"。

源数据集

• err_source_text：待修正的错误ASR-Diarization结果，包含dev和eval文件夹。
• ref_annotated_text：用于评估的人工标注参考文本，仅包含dev分割。

评估方法

• 使用MeetEval软件进行评估。
• 可通过命令行工具meeteval-wer cpwer评估修正后的SegLST文件。