A Hybrid Approach for Speech Enhancement with DNN Supported Acoustic Beamforming
Author
Term
4. term
Education
Publication year
2018
Submitted on
2018-06-07
Pages
80
Abstract
Modern hearing aids often have more than one microphone available for each device. It has been shown that substantial gains in speech intelligibility can to obtained by applying multichannel signal processing methods (e.g. beamformers) to noisy observations in noisy environments such as cocktail parties or restaurant-like environments. Model-based signal processing methods might, however, perform less well in acoustic environments where the SNR is low as the unknown parameters needed for the beamformers are harder to estimate. The motivation behind the work presented in this thesis, is thus to explore the possibility of applying a deep neural network (DNN) to support an acoustic beamformer as an alternative to the model-based methods. The DNN will in this thesis specifically estimate the direction-of-arrival (DOA) and the relative transfer function (RTF) vector needed for the examined beamformers. We have proposed three types of DNN supported beamformers in this thesis: 1) A minimum power distortionless response (MPDR) beamformer supported by a DNN for DOA estimation, 2) an MPDR beamformer supported by a DNN estimating RTF-vectors, and 3) a Bayesian beamformer where the posterior probabilities are estimated by a DNN. The experimental results show that the DNN-supported beamformers are able to outperform a model-based Bayesian beamformer in acoustic scenes with isotropic babble noise in terms of ESTOI, PESQ, and segSNR scores.
Keywords
Beamforming ; Machine Learning ; Deep Learning ; DNN ; Deep neural network ; Hearing aids ; MVDR ; signal processing ; statistical signal processing ; random processes ; RTF ; Bayesian beamformer ; Speech enhancement ; Noise reduction ; Array ; Microphone array ; DOA ; ESTOI ; PESQ ; acoustics ; HRTF ; maximum likelihood ; cross-entropy ; fourier transformation
Documents