Machine Learning Approaches for Intelligent Hearing Aid Control

• Forschungsstelle: INNOSUISSE
• Projektnummer: 8539.2;4 ESPP-ES
• Projekttitel Englisch: Machine Learning Approaches for Intelligent Hearing Aid Control

Kurzbeschreibung (Deutsch)

Machine Learning Approaches for Intelligent Hearing Aid Control

Kurzbeschreibung (Englisch)

Machine Learning Approaches for Intelligent Hearing Aid Control

Abstract (Deutsch)

This proposal addresses data analysis needs of Phonak Hearing Systems, i.e., to enhance hearing instruments signal processing or fitting processes with intelligent, autonomous decision making and adaptation capabilities. The research activities are concerned with A. classifier design for acoustic environments, B. online and reinforcement scenarios for intelligent hearing instrument control and C. source estimation for auditory scene analysis. The control problem is decomposed into a classification problem to characterize the acoustic environment and a mapping problem of acoustic classes to parameter settings of hearing instruments. The acoustic classes will be trained in a supervised way with automatic feature selection and model validation. Part of this task is concerned with the acquisition of representative training data and the design of a data base. Furthermore, unsupervised clustering methods for class discovery and weakly supervised reinforcement learning algorithms are employed to infer specific control parameters for hearing instruments which meet the hearing goals of hearing instrument users.

Abstract (Englisch)

This proposal addresses data analysis needs of Phonak Hearing Systems, i.e., to enhance hearing instruments signal processing or fitting processes with intelligent, autonomous decision making and adaptation capabilities. The research activities are concerned with A. classifier design for acoustic environments, B. online and reinforcement scenarios for intelligent hearing instrument control and C. source estimation for auditory scene analysis. The control problem is decomposed into a classification problem to characterize the acoustic environment and a mapping problem of acoustic classes to parameter settings of hearing instruments. The acoustic classes will be trained in a supervised way with automatic feature selection and model validation. Part of this task is concerned with the acquisition of representative training data and the design of a data base. Furthermore, unsupervised clustering methods for class discovery and weakly supervised reinforcement learning algorithms are employed to infer specific control parameters for hearing instruments which meet the hearing goals of hearing instrument users.