README.md

metadata

license: cc-by-4.0
task_categories:
  - audio-classification
language:
  - en
tags:
  - audio-quality
  - mos
  - ovl
  - xacle
pretty_name: OVL (Overall Audio Quality) MOS -- XACLE
dataset_info:
  - config_name: per_clip
    features:
      - name: audio_id
        dtype: string
      - name: wave_filename
        dtype: string
      - name: source
        dtype: string
      - name: split
        dtype: string
      - name: mean_score
        dtype: float32
      - name: std_score
        dtype: float32
      - name: median_score
        dtype: float32
      - name: n_ratings
        dtype: int32
      - name: batch
        dtype: string
    splits:
      - name: train
        num_examples: 100
    configs_note: per-clip aggregated MOS
  - config_name: per_rating
    features:
      - name: audio_id
        dtype: string
      - name: wave_filename
        dtype: string
      - name: source
        dtype: string
      - name: split
        dtype: string
      - name: score
        dtype: float32
      - name: listener_id
        dtype: string
      - name: batch
        dtype: string
    splits:
      - name: train
        num_examples: 460
configs:
  - config_name: per_clip
    data_files:
      - split: train
        path: per_clip/train-*
  - config_name: per_rating
    data_files:
      - split: train
        path: per_rating/train-*

OVL (Overall Audio Quality) MOS -- XACLE

Human MOS ratings of overall audio quality (OVL) for real XACLE-train audio clips, collected on Amazon Mechanical Turk. OVL here is synonymous with AQ (audio quality): annotators rate how natural and clean each clip sounds on a 1-5 scale, judging the audio itself -- not what it depicts, and without any text caption.

Each HIT presented 10 XACLE clips plus 1 hidden reference anchor (a known-good or known-bad clip) used only to screen annotators; anchors are excluded from this dataset. 4 low-quality assignments flagged on MTurk were rejected and dropped.

Configs

config	rows	one row per
per_clip	100	unique XACLE clip -- aggregated MOS
per_rating	460	single (assignment x clip) rating

per_clip

audio_id, wave_filename (XACLE-train file), source (xacle), split (train), mean_score (MOS), std_score, median_score, n_ratings, batch.

per_rating

audio_id, wave_filename, source, split, score (1-5), listener_id (anonymized worker), batch.

Audio is referenced by wave_filename (XACLE-train), not embedded.

Summary

• 100 unique XACLE-train clips, 460 ratings, 9 workers, ~4.6 ratings/clip.
• Mean OVL 3.19 (per-clip MOS mean 3.18).
• Inter-annotator agreement: Krippendorff alpha (interval) 0.388, ICC(1,1) 0.392, ICC(1,k) 0.748.
• Anchor probe: low anchors mean 2.04, high anchors mean 4.74.

source is xacle for every row; the column exists so RELATE OVL ratings can be appended to this repo later.