Modeling User Utterances as Intents in an Audiological Design Space

The global number of people living with hearing loss continues to grow, while the clinical resources are limited. To address this we describe a scalable goal oriented system. We outline a method on creating an audiological vocabulary, which can be mapped to intents. We create a shared audiological parameter space, with inspiration from clinical workflows. Matching of the intents and the audiological space, results in hearing aid fitting parameters, which then receive feedback from the user. We discuss how to train embedding and recurrent neural network models implementing attention mechanisms, to predict the optimal settings based on learned sequences of dialogue states and device fitting outcomes.