README.md

metadata

license: apache-2.0
task_categories:
  - audio-classification
language:
  - en
size_categories:
  - 1K<n<10K

Dataset Structure

This repo contains DailyTalkEdit and the extended semantic influence annotation of PartialEdit.

• daily_talk_edit/
  • concat_utts/: paired audio files per sample, obtained by concatenating all utterances in dialogues
    • <id>_original.wav
    • <id>_modified.wav
  • modified_utts/: modified-only audio segments
    • <id>_<utt_id>_modified.wav
  • metadata/
    • train.jsonl, val.jsonl, test.jsonl
      • filename: audio file name
      • fake_region: modified time ranges
      • semantic_influence: text description of semantic impact
    • modify_config_pipeline.json: modification details
• partial_edit/
  • metadata/semantic_influence.jsonl
    • utt_id: sample ID
    • influence: text description of semantic impact
• sampled_test_data/
  • *.txt: data IDs of the sampled data in the mixed evaluation set in the paper "HoliAntiSpoof: Audio LLM for Holistic Speech Anti-Spoofing".