An Open-Source Dataset for Infant Cry and Snoring Detection

Liu Q, Song L, Xu D, Long Y. An Open-Source Dataset for Infant Cry and Snoring Detection. J Audio Eng Soc. 2026;74(3):164-177. doi:10.17743/jaes.2022.0255

@article{Liu2026_23132,
  author = {Liu, Qingyu and Song, Longfei and Xu, Dongxing and Long, Yanhua},
  title = {{An Open-Source Dataset for Infant Cry and Snoring Detection}},
  journal = {Journal of the Audio Engineering Society},
  volume = {74},
  number = {3},
  pages = {164--177},
  year = {2026},
  month = mar,
  publisher = {Audio Engineering Society},
  doi = {10.17743/jaes.2022.0255},
  url = {https://doi.org/10.17743/jaes.2022.0255}
}

TY  - JOUR
TI  - An Open-Source Dataset for Infant Cry and Snoring Detection
AU  - Liu, Qingyu
AU  - Song, Longfei
AU  - Xu, Dongxing
AU  - Long, Yanhua
T2  - Journal of the Audio Engineering Society
J2  - J. Audio Eng. Soc.
VL  - 74
IS  - 3
SP  - 164
EP  - 177
PY  - 2026
DA  - 2026/03/06
DO  - 10.17743/jaes.2022.0255
UR  - https://doi.org/10.17743/jaes.2022.0255
PB  - Audio Engineering Society
LA  - en
AB  - The detection of infant cry and snoring events plays an important role in health monitoring and audio signal analysis. However, existing sound event detection datasets rarely provide sufficient, strongly labeled data for these specific sounds. This paper introduces the Infant Cry and Snoring Detection (ICSD) dataset, a novel open-source resource designed to advance research in this area. The ICSD comprises three subsets: a real strongly labeled set with manually annotated event boundaries, a weakly labeled set with clip-level annotations, and a synthetic set generated using the Scaper Toolkit with strong event-based labels. Furthermore, three baseline systems are established on the ICSD dataset, all built upon a convolutional recurrent neural network (CRNN) backbone: a mean-teacher CRNN system, a CRNN+bidirectional encoder representation from audio transformers (BEATs) system integrating pretrained BEATs audio representations, and a Competitive CRNN+BEATs system, which extends the previous model by incorporating speech, music, and noise categories to mitigate acoustic confusion. Experimental results show that the BEATs-based systems outperform the mean-teacher CRNN baseline, and the Competitive CRNN+BEATs model achieves superior robustness on real recordings with reduced false-positive rates. The ICSD dataset and proposed baselines are expected to serve as valuable open benchmarks for future infant cry and snoring detection research.
ER  -