Conversational AI Audiology: Remote, natural and automated testing of hearing and fitting

点击翻译按钮获取中文摘要

Conversational artificial intelligence (AI), i.e. technology that a patient can talk to, has the potential to remove barriers from access to healthcare and in particular hearing healthcare. Imagine that people who worry about their hearing would not need to visit a GP or an audiologist for a first diagnosis but could just ask their home assistant "Please check my hearing". This could lead to an increasing awareness of hearing loss and uptake of hearing aids by the 1.5 billion people (2.5 billion in 2050) who suffer from a hearing loss. It is estimated that at least 6 million people in the UK would benefit from a hearing aid but only 2 million have one, and uptake among them is low and slow. A hearing test that takes place as a conversation with a technical device has further advantages to conventional technologies: First, it tests real-world sounds (speech) rather than the audibility of faint synthetic sounds (pure tones in an audiogram). Second, a conversational AI system can directly simulate the speech that would be delivered through a hearing aid and thus quantify the benefit that the best hearing-aid setting would give. Third, remote testing does not only remove barriers in accessing healthcare but is also of benefit to vulnerable patients. In this project we will apply text-to-speech (TTS) and automatic speech recognition (ASR) to hearing tests in order to achieve these goals. TTS and ASR will allow the patient to interact easily with the test system, which is an additional benefit for those who struggle with graphical user interfaces, in particular some elderly patients. The communication between patient and AI will be driven by the goals to have a more natural interaction than in conventional tests and to characterise a hearing loss in short time. The candidate should have an interest in healthcare and is expected to have a background in deep learning and other fields of artificial intelligence, obtained from a first degree in engineering, computer science, mathematics, physics or similar. The candidate will acquire skills in TTS, ASR and active learning. Collaborations with other young researchers and activities like journal clubs will contribute to establishing a basic knowledge in audiology and clinical testing. The candidate will produce research outputs that are published in machine-learning conferences (NeurIPS, ICML, Interspeech, ICASSP) and leading hearing journals (Trends in Hearing, IJA, Hearing Research). Further references: Schlittenlacher, J., Baer, T. (2021) Text-to-speech for the hearing impaired, arXiv, 2012.02174