EARS: Expressive Anechoic Recordings of Speech

This is a mirror of the Expressive Anechoic Recordings of Speech (EARS) dataset. The original files were converted from WAV to Opus to reduce the size and accelerate streaming.

• Sampling rate: 48 kHz
• Channels: 1
• Format: Opus
• Splits:
  • Train: 92 hours, 15939 utterances, speakers p001 to p099
  • Validation: 2 hours, 322 utterances, speakers p100 and p101
  • Test: 6 hours, 966 utterances, speakers p102 to p107
• License: CC BY-NC 4.0
• Source: github.com/facebookresearch/ears_dataset
• Paper: EARS: An Anechoic Fullband Speech Dataset Benchmarked for Speech Enhancement and Dereverberation

Usage

import io

import soundfile as sf
from datasets import Features, Value, load_dataset

for item in load_dataset(
    "philgzl/ears",
    split="train",
    streaming=True,
    features=Features({"audio": Value("binary"), "name": Value("string")}),
):
    print(item["name"])
    buffer = io.BytesIO(item["audio"])
    x, fs = sf.read(buffer)
    # do stuff...

Citation

@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 = {Proc. Interspeech},
  pages = {4873--4877},
  year = {2024},
}