README.md

metadata

language:
  - zh
  - en
task_categories:
  - audio-to-audio
tags:
  - emotional-intelligence
  - speech-dialogue
  - audio
size_categories:
  - 1K<n<10K
license: cc-by-4.0

EmoS Test Sets

This repository is a test-only release of the EmoS speech-dialogue data. It contains no training audio or training metadata.

Contents

• archives/test/cn.tar and archives/test/en.tar: bilingual standard evaluation data with 1,020 scenarios and 4,080 scored model responses.
• archives/test50_human/cn.tar and archives/test50_human/en.tar: blinded human-evaluation packages for the same 1,020 scenarios and 4,080 WAV files.

All TAR archives are uncompressed and preserve their documented internal directory layout. Download a package and extract it with tar -xf <archive>.tar. ARCHIVES.tsv lists each package SHA-256 checksum.

Neither test surface may be used for training.

License

This dataset is released under the CC BY 4.0 license.