Pre-processing techniques for improved detection of vocalization sounds in a neonatal intensive care unit

• DOI: 10.1016/j.bspc.2017.07.024
• Publication Date: 2017-08-30T11:01:15Z
• Authors: Ganna Raboshchuk; Climent Nadeu; Sergio Vidiella P...; Oriol Ros Fornells; Blanca Muñoz Mahamud; Ana Riverola de V...
• Keywords: Discriminative model

The sounds occurring in the noisy acoustical environment of a Neonatal Intensive Care Unit (NICU) are thought to affect the growth and neurodevelopment of preterm infants. Automatic sound detection in a NICU is a novel and challenging problem, and it is an essential step in the investigation of how preterm infants react to auditory stimuli of the NICU environment. In this paper, we present our work on an automatic system for detection of vocalization sounds, which are extensively present in NICUs. The proposed system reduces the presence of irrelevant sounds prior to detection. Several pre-processing techniques are compared, which are based on either spectral subtraction or non-negative matrix factorization, or a combination of both. The vocalization sounds are detected from the enhanced audio signal using either generative or discriminative classification models. An audio database acquired in a real-world NICU environment is used to assess the performance of the detection system in terms of frame-level missing and false alarm rates. The inclusion of the enhancement pre-processing step leads to up to 17.54% relative improvement over the baseline. Peer Reviewed