TRACE — Temporal Relationship-Aware Conversational Entrainment Detection in Dyadic Speech

TRACE is a binary naturalness-labeled dataset of dyadic spoken interactions, designed for training and evaluating speech naturalness classifiers. Each sample pairs two speaker audio tracks with metadata about their relationship and conversational context.

The dataset contains 8,434 labeled dyadic pairs across four augmentation conditions, split into train (5,131) and test (3,303) sets.

Paper: arxiv.org/abs/2606.30543

---

Augmentation types

augmentation_type	Description
original	Unmodified naturalistic speech from the Seamless Interaction corpus
original_vc	Speaker-A and Speaker-B tracks individually voice-converted via SeedVC
emotivoice_tts	Speaker-B replaced with EmotiVoice TTS conditioned on contrastive emotion
emotivoice_tts_vc	Same as emotivoice_tts, with Speaker-B TTS additionally voice-converted

naturalness = 1 for original pairs; naturalness = 0 for all augmented conditions.

---

Dataset structure

TRACE/
├── train.csv               # 5,131 rows with updated audio paths
├── test.csv                # 3,303 rows with updated audio paths
├── missing_files.txt       # 169 source files absent at collection time
└── audio/
    ├── original/           # naturalistic WAVs (from Seamless corpus)
    ├── original_vc/        # voice-converted naturalistic WAVs
    ├── emotivoice_tts/     # TTS-resynthesised Speaker-B WAVs + mixed track
    └── emotivoice_tts_vc/  # TTS + VC Speaker-B WAVs + mixed track

CSV columns

Column	Description
row_id	Unique row identifier
augmentation_type	See table above
naturalness	Binary label: 1 = natural, 0 = unnatural
high_level_context	Conversational context category
participant1_relpath_abs	Absolute path to Speaker-A audio
participant2_relpath_abs	Absolute path to Speaker-B audio
rel_detail	Speaker relationship description
source_row_p1	Source row ID for Speaker A
source_row_p2	Source row ID for Speaker B
mixed_relpath_abs	Path to mixed stereo track (TTS conditions only)
transcript	Transcript text (where available)

---

Source data and attribution

• Seamless Interaction corpus — Meta AI. Naturalistic dyadic speech recordings. CC BY-NC 4.0
• EmotiVoice — NetEase Youdao. Emotion-conditioned TTS. Apache 2.0
• SeedVC — Voice conversion model used for speaker identity transfer.

---

License

This dataset is released under CC BY-NC 4.0 in accordance with the Seamless Interaction corpus license. Non-commercial use only.

---

Citation

If you use this dataset, the pretrained weights, or the associated code in your research, please cite:

@misc{ugandhar2026tracetemporalrelationshipawareconversational,
      title={TRACE: Temporal Relationship-Aware Conversational Entrainment Detection in Dyadic Speech}, 
      author={Sathvik Manikantan Napa Ugandhar and Hao Zhang and Alison Gunzler and Yuzhe Wang and Thomas Thebaud and Georgi Tinchev and Venkatesh Ravichandran and Laureano Moro-Velázquez},
      year={2026},
      eprint={2606.30543},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2606.30543}, 
}