Detecting hearing loss from children’s speech using machine learning

Jessica Monaghan, David Allen

National Acoustic Laboratories, Sydney, Australia

Background: Undiagnosed hearing loss in children can have serious consequences for the development of speech and language skills, ultimately impacting educational and social outcomes as well as quality of life. In many countries, including Australia, newborn screening for hearing loss is common or even mandated. However, over 50% of children diagnosed with hearing loss are identified subsequent to newborn hearing screening, either because their hearing loss was too mild to detect at birth or developed later. Since no widespread screening of preschool or school-aged children exists, hearing loss during the early years may go undetected. In addition to sensorineural hearing loss, preventable and potentially reversible hearing loss in the form of ‘glue’ ear affects 1 in 5 preschool children at any given time, and is especially prevalent in Aboriginal and Torres Strait Islander children. Evidence indicates parents are not able to identify reliably even severe hearing loss in their children, so relying on their suspicions of hearing impairment is not sufficient to address the problem of undiagnosed childhood hearing loss. Testing preschool children, particularly in the age range 2-3 years, is time consuming and requires specialist equipment and expertise, making large-scale testing unfeasible.

Results & Conclusion: Although the trajectory of speech development varies widely across children, hearing loss will nevertheless affect the perception and, therefore, the production, of different phonemes differently, depending on spectral frequency. Other characteristics such as pitch or rate of speech may also be altered by hearing impairment. These differences may be too subtle for humans to observe, but our preliminary analysis indicates that deep-learning techniques can be used to differentiate the speech of children with and without hearing loss. This approach has the potential to be used as a convenient screening tool for children or to monitor speech development in children at risk of otitis media.

Monaghan et al
Here we explore the potential for using machine learning to detect hearing loss from children’s speech.

insert video