Submitted by: Shane Lani, Jennifer Cooper, T.J. Flynn,Adaleena Mookerjee,Jordan Schleif,Olivia Ott
New parents want to keep an ear on their new little ones while they’re sleeping. An audio baby monitor, consisting of a baby monitor unit near the child’s bed and a portable parental unit, allows them to walk around the house or socialize with the neighbors, while at the same time having an ear out and keeping tabs on the newborn. The monitor alerts the parents when their bundle of joy is upset, so they can go over to interpret whether the cries indicate a need for food, for a diaper change or just for an expression of general discontent.
Most off-the-shelf baby monitors however also alert parents when irrelevant sounds occur: noisy machines like an air conditioning, lullaby tunes, as well as transient passing motorcycles and overpassing airplanes. Any sound above a certain threshold will trigger the portable parental unit, giving false alarms (appropriately called ‘nuisance alarms’).
The research team, three of whom welcomed a new baby into their family in the 6 months before the ASA challenge, believe that a smarter and better baby monitor is possible. It should distinguish a baby-cry from other loud sounds. They submitted the first steps for the development and testing of such a smart baby monitor, using Tympan as an open-source development platform. The idea is that the Tympan, being able to process signals in real time, analyzes the incoming sound and only passes it through when the algorithm has identified it as cries from a baby.
To distinguish baby cries from other sounds, the researchers identified a number of typical characteristics: the fundamental harmonic, which typically lies between 400-600 Hz, the harmonic structure and the temporal variations in the amplitude of the sound. The required algorithms use frequency domain processing and ‘cepstral’ or ‘quefrency domain’ processing. ‘Cepstral’ is a muddled derivative of the word ‘spectral’, and ‘quefrency’ a muddled derivative of ‘frequency’. Aptly named, because the method uses a derivative of the signal spectrum: the inverse Fourier transform of the logarithm. Cepstral processing is particularly useful for distinguishing echos, reflections and overtones in a signal. In this case study, it proved to be quite suitable for distinguishing babies cries from household noise such as aircraft and traffic noise.
These baby-steps in the development of a smart baby monitor have yielded some useful insights. Firstly, the Tympan Rev-E has proved itself as a suitable development platform for the proof-of-concept for a ‘smart’ baby monitor. Secondly: cepstral processing is quite effective for distinguishing baby-like sounds from non-baby-like sounds. However, human speech can still be easily misclassified as a crying baby, particularly when it comes to children’s voices. Some interesting ideas were put forward by the researchers to improve this: time domain processing (looking for rhythms, like hiccoughs), looking at multiple metrics simultaneously and applying machine learning techniques, already successfully used in image classification techniques, to audio applications.
For more information about the author please visit https://ep.jhu.edu/faculty/shane-lani.
About the TYMPAN ASA2021 Design Challenge
During the ASA (Acoustical Society of America) conference in June 2021, Tympan hosted a design challenge: What is possible with the Tympan?
10 exciting new applications were submitted and presented at the following ASA conference: Enhancements of hearing aids, spatial acoustic processing and smart earphones and much more. Stay tuned if you want to learn what is possible and to keep track of future developments.