hhoangphuoc/ami-av

Dataset Summary

This is the processed Audio-Visual Dataset from AMI Meeting Corpus. The dataset was segmented into sentence-level audio/video segments based on the individual [meeting_id]-[speaker_id] transcripts. The purpose of this data is for audio-visual speech recognition task (AVSR), particularly for spontaneous conversational speech.

General information about dataset:

Total #segments: 83,438 (including either audio/video or both)

Dataset({
    features: ['id', 'meeting_id', 'speaker_id', 'start_time', 'end_time', 'duration', 'transcript', 'audio', 'video', 'has_audio', 'has_video', 'has_lip_video', 'has_transcript'],
    num_rows: 83438
})

in which:

• #audio: 80,285 ~ 13GB
• #videos: 78,685 items ~ 5,6GB
• #lip videos: 67,438 items ~ 1.9GB

Additional Remarks: Audio are segmented and resampled to 16kHz, .wav format; Videos are resampled to 25fps, in .mp4 format

Dataset Folder Structure

Each file of the data have the unique segment_id in the format:

  [meeting_id]-[speaker_id]-[start_time]-[end_time]-[audio/video/lip_video].(wav/mp4)

The original folder structure that used to store this information is followed:

ami /
  |_ audio_segments/
      |_ ES2001a-0.00-0.10-audio.wav
      |_ ...
  |_ video_segments /
      |_ original_videos
          |_ ES2001a-0.00-0.10-video.mp4
          |_ ...
      |_ lips
          |_ ES2001a-0.00-0.10-lip_video.mp4
          |_ ...

Create this dataset

To replicate the creation of this dataset, you can download the original AMI Meeting Corpus. This dataset processed from all meeting recordings, original video is "Low-size DivX AVI videos" option and original audio is "Individual headsets" option.

The details steps to preprocess from original AMI corpus can be followed by this guidelines: https://github.com/hhoangphuoc/AVSL/blob/main/docs/Preprocess.md