Hier finden Sie wissenschaftliche Publikationen aus den Fraunhofer-Instituten.

A 2-stage approach for joint noise reduction and dereverberation by means of multi-channel equalization and a noise processor

: Sharma, A.; Goetze, S.

Kollmeier, Birger ; Deutsche Gesellschaft für Akustik -DEGA-, Berlin; Deutsche Gesellschaft für Audiologie -DGA-:
Akustik & Audiologie "Hören für alle". CD-ROM : DAGA 2014, 40. Jahrestagung für Akustik "Fortschritte der Akustik", 10.-13. März 2014, Oldenburg; DGA , 17. Jahrestagung "Akustik & Audiologie - Hören für alle", 12.-15. März 2014, Oldenburg; Tagungs-CD
Oldenburg: DEGA, 2014
ISBN: 978-3-939296-06-5
Deutsche Jahrestagung für Akustik (DAGA) <40, 2014, Oldenburg>
Deutsche Gesellschaft für Audiologie (DGA Jahrestagung) <17, 2014, Oldenburg>
Fraunhofer IDMT ()

Noise reduction (NR) for audio signals has been extensively studied during the last decades. Various approaches ranging from single channel, multi-channel and hybrid NR techniques have been proposed in the literature. Within the last years, signal enhancement for dereverberation received much attention. The main challenge in dereverberation by means of robust acoustic channel equalization has been the problem of identifying the room impulse response (RIR) accurately. In 2013, Rajan and Andy proposed acoustic multi-channel equalization multiple filters (MCEQ-MF) for imperfect RIR measurements. They suggested a system delay based on minimum $l 2$ norm to design equalization filters. This paper shows that incorporating regularization in MCEQ-MF algorithm enhances equalization results. In this paper, joint NR and dereverberation combining regularized MCEQ-MF for dereverberation and hybrid NR based on sub-band filtering is proposed. Individual microphone signals are first equalized by designing the equalizer filters based on regularized MCEQ-MF algorithm. Then, the equalized signals are post-processed using hybrid filter for noise reduction. Although noise part may be amplified after the equalization step, our results show that the combined approach is able to achieve dereverberation to a great extent even in presence of channel estimation errors as well as noise reduction after post-filtering.