DEAR

Dataset Summary

The Deep Evaluation of Audio Representations (DEAR) dataset is a benchmark designed to assess general-purpose audio foundation models on properties critical for hearable devices. It comprises 1,158 mono audio tracks (30 s each), spatially mixing proprietary anechoic speech monologues with high-quality everyday acoustic scene recordings from the HOA‑SSR library. DEAR enables controlled evaluation of:

• Context (environment type: domestic, leisure, nature, professional, transport; indoor/outdoor; stationary/transient noise)
• Speech sources (speech presence detection; speaker count)
• Acoustic properties (direct-to-reverberant ratio DRR, reverberation time RT60, signal‑to‑noise ratio SNR)

All tracks are down‑mixed to a single channel at 44.1 kHz (32‑bit) and split into development and test sets with no overlap in speakers, backgrounds, or impulse responses.

Tasks

Dataset Structure

├── data/
│   ├── 00094903-4dbf-44a9-bf09-698fc361dbff.wav
│   └── …
├── development.csv
└── test.csv

• .wav files: mono, 44.1 kHz, 32‑bit float
• .csv files: meta-data for all tasks, linkable to wav files with id

Usage

Visit the dedicated code repository: https://github.com/DEAR-dataset/code

Source Data

• Speech monologues (proprietary anechoic recordings)
• HOA‑SSR library ambisonics scenes (licensed via FORCE Technology)
• Impulse responses for controlled reverberation

Citation

If you use DEAR in your research, please cite:

@inproceedings{
  groeger2025dear,
  author={Gröger, Fabian and Baumann, Pascal and Amruthalingam, Ludovic and Simon, Laurent and Giurda, Ruksana and Lionetti, Simone},
  booktitle={ICASSP 2025 - 2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)}, 
  title={Evaluation of Deep Audio Representations for Hearables}, 
  year={2025},
  doi={10.1109/ICASSP49660.2025.10887737}
}

ArXiv version: arxiv.org/abs/2502.06664