README.md

---
language:
- en
license: apache-2.0
size_categories:
- 1K<n<10K
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
dataset_info:
  features:
  - name: meeting_id
    dtype: string
  - name: label
    dtype: int64
  - name: start_time
    dtype: float64
  - name: end_time
    dtype: float64
  - name: transcript_without_speaker
    dtype: string
  - name: valid_speakers
    sequence: string
  - name: speaker_order
    sequence: string
  - name: num_speakers
    dtype: int64
  - name: audio_data
    dtype: string
  - name: sample_rate
    dtype: int64
  splits:
  - name: train
    num_bytes: 10963079362
    num_examples: 2955
  download_size: 9325667972
  dataset_size: 10963079362
---
# CATS-ami-speaker-diarization Dataset

## Overview

This dataset is designed for speaker diarization tasks on the CATS (Comprehensive Assesment for Testing Speech) Dataset. It contains audio segments from the AMI Meeting Corpus with corresponding transcriptions and speaker information.

## Dataset Structure

Each example in the dataset contains:

- **meeting_id**: Identifier for the source meeting
- **label**: Segment identifier within the meeting
- **start_time/end_time**: Timestamp boundaries for the audio segment
- **transcript_without_speaker**: Text transcription without speaker attribution
- **valid_speakers**: List of unique speakers in the segment
- **speaker_order**: Sequential list of speakers for each sentence
- **num_speakers**: Count of unique speakers in the segment
- **audio_data**: Base64-encoded WAV audio data
- **sample_rate**: Audio sample rate in Hz

## Usage

This dataset is specifically designed for speaker diarization tasks, which involve determining "who spoke when" in multi-speaker audio recordings.

### Loading the Dataset

```python
from datasets import load_dataset

# Load the dataset
dataset = load_dataset("rma9248/CATS-ami-speaker-diarization")

# Access an example
example = dataset["train"][0]

# Get metadata
meeting_id = example["meeting_id"]
transcript = example["transcript_without_speaker"]
```

### Working with the Audio

```python
import base64
import io
from IPython.display import Audio

# Decode base64 audio data
audio_bytes = base64.b64decode(example["audio_data"])

# For playback in Jupyter notebooks
Audio(audio_bytes, rate=example["sample_rate"])

# Save to file
with open("sample_audio.wav", "wb") as f:
    f.write(audio_bytes)
```

### Diarization Task Example

The main task is to assign speaker labels to each part of the transcript based on the audio, please see our git repo when it's available

## Dataset Creation

This dataset was created from the AMI Meeting Corpus, focusing on segments with clear speaker turns. Each segment contains approximately 20 sentences from the original meetings, with audio segments extracted to match these sentence boundaries.

## Citation

If you use this dataset in your research, please cite the original AMI Meeting Corpus for now:

```
@inproceedings{mccowan2005ami,
  title={The AMI meeting corpus},
  author={McCowan, Iain and Carletta, Jean and Kraaij, Wessel and Ashby, Simone and Bourban, S and Flynn, M and Guillemot, M and Hain, T and Kadlec, J and Karaiskos, V and others},
  booktitle={International Conference on Methods and Techniques in Behavioral Research},
  volume={88},
  pages={100},
  year={2005}
}
```

## License

This dataset follows the licensing terms of the AMI Meeting Corpus, which is available for research and education purposes.