NathanRoll/ears-naturalistic-phrases

EARS naturalistic phrases (2-5 s)

6,153 short naturalistic English phrases (2-5 seconds, median 3.5 s) cut at pause boundaries from the freeform (unscripted) portions of the EARS dataset (Richter et al., Interspeech 2024): the emotional freeform monologues and the unconstrained freeform monologues. Reading tasks, interjections, and non-verbal recordings were excluded.

• Audio: 48 kHz, mono, 16-bit FLAC (lossless from the source WAV)
• Speakers: 98 (EARS train speakers p001-p099)
• Recording: anechoic chamber, studio-grade full-band speech
• Segmentation: energy-based phrase detection (25 ms frames, 250 ms pause threshold), 120 ms context padding, 10 ms edge fades; segments were QA filtered for level (RMS >= -38 dBFS), clipping (< 0.1% samples), and speech density (>= 55% voiced frames)

License: CC BY-NC 4.0, same as the source dataset. If you use this data, cite the EARS paper:

@inproceedings{richter2024ears,
  title={{EARS}: An Anechoic Fullband Speech Dataset Benchmarked for Speech Enhancement and Dereverberation},
  author={Richter, Julius and Wu, Yi-Chiao and Krenn, Steven and Welker, Simon and Lay, Bunlong and Watanabe, Shinjii and Richard, Alexander and Gerkmann, Timo},
  booktitle={ISCA Interspeech},
  year={2024}
}