• Julius Garde
4. semester, Mathematical Engineering, Master (Master Programme)
In this thesis, we propose a speech enhancement system for own-voice retrieval the presence of additive background. The system is designed with embedded devices in mind, specifically hearing assistive devices. The work focuses on discriminating noise-dominated time-frequency units from own-voice-dominated units, for which we employ a convolutional neural network to perform classification. Using only the classified time-frequency units, we show that a MVDR and a MWF beamformer can be constructed. Our results show that considerable improvements can be made in terms of perceived quality and intelligibility using the MVDR beamformer for selected noise types, whereas speech-shaped noise and babble-noise remains a challenge.
Publication date7 Jun 2019
External collaboratorOticon Danmark AS
Jesper Jensen jesj@demant.com
ID: 305337391