PhD student: Mathias Bach Pedersen
Supervisors: Søren Holdt Jensen, Asger H. Andersen and Jesper Jensen
Support: Independent Research Fund Denmark
Digital processing of speech is broadly utilized in modern technology, from mobile phones to hearing aids. Whenever digitally processed speech is presented to a human listener, the intelligibility, i.e. the average percentage of understood words or syllables of the speech, should be as high as possible. To design speech processing technology for optimizing intelligibility, the ability to easily and reliably measure or predict the intelligibility of a speech signal under any given noise, interference or processing is greatly desirable. There are many factors that play into the intelligibility of a speech signal, including the type and intensity of noise, the type of processing, whether the speech is presented to both ears of the listener, and whether the listener has any hearing impairment. Much research has already been done on monaural intelligibility prediction, under diverse noise conditions and state-of-the-art speech processing methods. More recent research has focused on generalizing existing methods to binaural intelligibility prediction, or to predict intelligibility non-intrusively, i.e. without the use of a clean reference signal. In this project the aim is to develop a simultaneously binaural and non-intrusive intelligibility predictor. These criteria are motivated by a desire to integrate intelligibility prediction into binaural hearing assistive devices, where a clean reference signal is not available. This predictor should be accurate for as wide a range of noise and processing conditions as possible. To achieve this the predictor will be designed using deep learning techniques, motivated by the recent success in related fields such as speech enhancement.