AMI-universal

Description

The dataset is called universal because it can be used the evaluate ASR, VAD and Speaker Diarization.

Dataset contains the following:

• audio
• segment information
• speaker information
• transcripts

The dataset is a combination of two other datasets to include information for all three tasks.

Original datasets come from here:

• https://huggingface.co/datasets/diarizers-community/ami
• https://huggingface.co/datasets/diarizers-community/ami_for_diarizationlm

Subsets:

From the diarizers-community/ami dataset I only took the sdm (microphone recording) subset, which is equivalent to the diarizers-community/ami_for_diarizationlm dataset.

Included splits:

• validation split
• test split

The train split was not consistent in both datasets and I assume we don't need it either.

dataset_info:
  features:
  - name: audio
    dtype:
      audio:
        sampling_rate: 16000
  - name: timestamps_start
    list: float64
  - name: timestamps_end
    list: float64
  - name: speakers
    list: string
  - name: transcript
    list: string
  splits:
  - name: validation
    num_bytes: 1114378839
    num_examples: 18
  - name: test
    num_bytes: 1044618289
    num_examples: 16
  download_size: 2035676988
  dataset_size: 2158997128
configs:
- config_name: default
  data_files:
  - split: validation
    path: data/validation-*
  - split: test
    path: data/test-*