Postdoc Researcher: Morten Kolbæk
Supervisors: Søren Holdt Jensen, Zheng-Hua Tan and Jesper Jensen
Support: Independent Research Fund Denmark
Speech intelligibility enhancement is a goal that has been pursued for a long time. Generally, it is hard to achieve speech intelligibility improvements for normal hearing with single-microphone methods, in particular if no special knowledge is available about the target speaker and/or the noise background. For hearing impaired subjects, on the other hand, single-microphone methods can improve intelligibility quite significantly. With multi-microphone methods – which allow for spatial filtering approaches / beamforming – speech intelligibility enhancement is an easier task and improvements are possible also for normal-hearing. Recent years have seen progress in the areas of speech intelligibility models, binaural beamforming and deep learning for speech enhancement.
The basic goal of this project is to combine these advances to build binaural, multi-microphone, speech intelligibility enhancement systems for normal-hearing and for hearing impaired that perform optimally with respect to state-of-the-art speech intelligibility predictors.
We envision a perceptually guided speech enhancement framework which – based on multiple noisy microphone signals recorded on either ear of the user of a hearing assistive device – can process the noisy signals using deep neural networks to produce two enhanced output signals, a left- and a right-ear signal, of improved intelligibility.