hhoangphuoc/ami-disfluency

DSFL Dataset - AMI Disfluency Laughter Events

This dataset contains segmented audio and video clips from AMI Meeting Corpus, which only consisted of disfluencies and laughter events, segmented in both audio and visual modality.

This dataset, along with hhoangphuoc/ami-av is created for my research related to Audio-Visual Speech Recognition, which I currently developed at: https://github.com/hhoangphuoc/AVSL

For reproducing the work I've done to create this dataset, checkout the documentations: https://github.com/hhoangphuoc/AVSL/blob/main/docs/Preprocess.md

Summary of the Dataset

• Number of recordings: 35,731
• Has audio: True
• Has video: True
• Has lip video: True

Dataset({
    features: ['id', 'meeting_id', 'speaker_id', 'start_time', 'end_time', 'duration', 'has_audio', 'has_video', 'has_lip_video', 'disfluency_type', 'transcript', 'audio', 'video', 'lip_video'],
    num_rows: 35731
})

The dataset contains ~35k recordings of disfluency and laughter event in AMI Corpus. However, not all the recording contains all audio/video sources. In details, the number of recordings corresponding to each sources are:

• #audio: 35,731 items
• #video: 35,536 items
• #lip_videos: ~28,263 items

Specific to the disfluency and laughter events:

• #laughter: Total 469 laughs events, all have audio sources, 450/469 have video sources while only 346 items have lip_video
• #disfluency: Contains 10 types of disfluencies (annotated in AMI transcripts). All have disfluent words and audio available, however, #video and #lip_video are varied.

Using the dataset

The overall information of the dataset have been reported in dsfl-segments-info.csv files. This including:

• id: unique segment_id of the event
• meeting_id: the meeting where this event belong to
• speaker_id: the corresponding speaker of such events
• start_time: start timestamp of the event
• end_time: stop timestamp of the event
• disfluency_type: type of disfluency corresponding to that event. Type is laugh if it is laughter
• transcript: the corresponding disfluent word. Annotated as <laugh> if it is laughter event.
• audio | video | lip_video : Path to the corresponding sources

Similarly, metadata.jsonl file contains the information of each recordings items in JSON line format, i.e. each line refered to each segment_id.

To use the dataset, following these steps:

1. Download the dataset using load_dataset("hhoangphuoc/ami-disfluency")

2. Alternatively, if load dataset doesn't include the audio/video recording resources. Download the following files manually:

   • audio_segments.tar.gz: Including all audio segments
   • video_segments.tar.gz: Including both original_video (in original_video folder) and the lips video (lips folder)

   or you can do it by using wget:

   wget https://huggingface.co/datasets/hhoangphuoc/ami-disfluency/resolve/main/audio_segments.tar.gz
   
   wget https://huggingface.co/datasets/hhoangphuoc/ami-disfluency/resolve/main/video_segments.tar.gz
   

The folder structure where you can store the data can be as following example:

  path/to/folder(ami-disfluency) /
      |_ dsfl-segments-info.csv
      |_ audio_segments
          |_ ES2001a-0.00-0.01-audio.wav
          |_ ...
      |_ video_segments
          |_ original_videos
                |_ ES2001a-0.00-0.01-laugh-video.mp4
                |_ ...
          |_ lips
                |_ ES2001a-0.00-0.01-laugh-lip_video.mp4
                |_ ...
      |_ ...