| Popis: |
This thesis is focused on the application of computer audition (i. e., machine listening) methodologies for monitoring states of emotional wellbeing. Computer audition is a growing field and has been successfully applied to an array of use cases in recent years. There are several advantages to audio-based computational analysis; for example, audio can be recorded non-invasively, stored economically, and can capture rich information on happenings in a given environment, e. g., human behaviour. With this in mind, maintaining emotional wellbeing is a challenge for humans and emotion-altering conditions, including stress and anxiety, have become increasingly common in recent years. Such conditions manifest in the body, inherently changing how we express ourselves. Research shows these alterations are perceivable within vocalisation, suggesting that speech-based audio monitoring may be valuable for developing artificially intelligent systems that target improved wellbeing. Furthermore, computer audition applies machine learning and other computational techniques to audio understanding, and so by combining computer audition with applications in the domain of computational paralinguistics and emotional wellbeing, this research concerns the broader field of empathy for Artificial Intelligence (AI). To this end, speech-based audio modelling that incorporates and understands paralinguistic wellbeing-related states may be a vital cornerstone for improving the degree of empathy that an artificial intelligence has. To summarise, this thesis investigates the extent to which speech-based computer audition methodologies can be utilised to understand human emotional wellbeing. A fundamental background on the fields in question as they pertain to emotional wellbeing is first presented, followed by an outline of the applied audio-based methodologies. Next, detail is provided for several machine learning experiments focused on emotional wellbeing applications, including analysis and recognition of under-researched phenomena in speech, e. g., anxiety, and markers of stress. Core contributions from this thesis include the collection of several related datasets, hybrid fusion strategies for an emotional gold standard, novel machine learning strategies for data interpretation, and an in-depth acoustic-based computational evaluation of several human states. All of these contributions focus on ascertaining the advantage of audio in the context of modelling emotional wellbeing. Given the sensitive nature of human wellbeing, the ethical implications involved with developing and applying such systems are discussed throughout. |
