Sat | Sep 22, 2018

Artificial Intelligence and the Economy | Tackling hearing loss

Published:Monday | June 11, 2018 | 12:00 AMJordan Micah Bennett/Contributor

Artificial Intelligence and the Economy features machine learning computer models in Jamaica. These models are computer algorithms, or smart apps, that seek to give computers the ability to learn like children for a variety of tasks. Here, we highlight how an author's work may solve a particular set of real-world tasks or problems. By doing this, we aim to foster more and more machine, learning works, to be done by more and more Jamaican people.

 

Focus

 

Today, we'll highlight the machine-learning work, a paper/algorithm called 'Modelling Sensorineural Hearing-impaired Listeners' Perception of Speaker Intelligibility in Noise", by UWI lecturers Dr Lindon W. Falconer, Dr AndrÈ Coy, and their overseas colleague, Professor Jon Barker.

Jordan: How would you describe your work?

Dr Coy, et al: Disabling hearing loss is a major challenge faced by many individuals in societies throughout the world. The World Health Organization (WHO) has reported that approximately 6.1 per cent of the world's population has disabling hearing loss, and about 93 per cent of these people are adults. The percentages of persons affected are higher in developing countries than developed countries.

The WHO defines disabling hearing loss as a permanent hearing loss of 40 dB in the better ear of an adult and 30 dB in the better hearing of a child. Persons with disabling hearing loss have problems communicating verbally, especially in noisy environments.

Many researchers have shown that people with disabling hearing loss are more likely to suffer from anxiety, depression, and a reduced social life. This deficit may cause individuals to lose their jobs or reduce their earning potential. Therefore, disabling hearing loss is affecting the economic well-being of families, communities, and countries.

The hearing aid is the main rehabilitative device used to improve hearing for individuals with disabling hearing loss. However, despite the advancement made in hearing-aid technology, users still complain about the unsatisfactory performance of the device in noisy environments such as a bus stop, a crowded cafeteria, or a sporting event. Many hearing-aid users, therefore, avoid these locations or stop wearing the hearing aid. Therefore, the rehabilitative effect the hearing aid aims to achieve is not fully met; hence improvement in hearing-aid algorithms is needed.

Developing and testing new hearing-aid algorithms is faced with challenges when testing on human subjects to verify the algorithm performance. These challenges include the time and expense to locate listeners with specific hearing loss and conduct listening tests. Some listeners may take weeks or months to adjust to the "new" sound produced by the new algorithm, therefore, the performance verification time can be lengthy. A solution to these challenges is to replace the human listener with an artificially intelligent computational model (AICM) that mimics the hearing loss of an individual and predicts the intelligibility of speech and speakers that the person perceives in noise. If the AICM performs close to human listeners, then it can be used to inform the development of an application that will aid in the rapid development and testing of hearing-aid algorithms.

The AICM was developed as a computer programme using signal processing techniques to mimic hearing loss and machine-learning algorithms to predict the intelligibility of speech. It was configured with the audiograms of the hearing losses to be simulated and trained using audio recordings of hundreds of words produced by male and female speakers.

To test the performance of the AICM, experiments were conducted on normal hearing human listeners with simulated hearing impairments and the AICM at the University of Sheffield, England, and the University of the West Indies, Jamaica. Comparison of the results between the AICM and human listeners showed that the AICM mimics human listeners closely in areas such as average speech recognition scores, average digits recognition scores, and recognition scores of female speakers. The AICM didn't perform as well on average recognition scores of letters and male speakers. Further experiments will be done to tune the performance of the AICM and test its ability to predict the performance of some common hearing-aid algorithms.

 

Jordan: What do you think is the short-term impact of your paper's algorithm/model on Jamaican society?

Dr Coy, et al: The immediate impact this work will have on the Jamaican society is to show that we can conduct successful research in machine learning and artificial intelligence at the highest level.

The research is being done with the aim of informing developers of hearing-aid algorithms about our model to predict how users of hearing aids will rate the intelligibility of speech and speakers when the hearing aids are used in noise.

Jordan: Also, what is the long-term impact that you envision for your paper's algorithm/model on Jamaican society?"

Dr Coy, et al: If the work, or a related approach is ultimately successful in predicting intelligibility, Jamaican hearing-aid users would benefit from more accurate and targeted fitting, a shorter adjustment period, and a far better overall user experience.

Jordan: Great! What are some shortfalls of the algorithm?

Dr Coy et al: Sensorineural hearing loss can be associated with audibility and suprathreshold losses. Although audibility loss is the major contributor to most sensorineural hearing impairments, suprathreshold loss can be significant, especially in moderate to severe hearing impairment. The method used to simulate hearing impairment in the algorithm only took audibility loss into consideration.

Jordan: What methods could be used to improve the learning algorithms?

Dr Coy, et al: A method that can be used to improve the performance of the machine learning algorithm used in the research is to use deep learning neural networks (DNN) as the machine-learning algorithm instead of Hidden Markov Models and Gaussian Mixture Models.

Jordan: Doctors, do you plan to work on any more machine learning stuff in the future?

Dr Falconer: I plan to use machine-learning algorithm in the implementation of monitoring devices that can be used for safety and security purposes.

Jordan: Doctors, thanks a lot and I look forward to your future work.

Stay tuned for next week, when we'll speak with more machine learning-aligned authors.

Send feedback to editorial@gleanerjm.com, or jordanmicahbennett@gmail.com.