Multi-channel linear prediction-based speech dereverberation with low-rank power spectrogram approximation

Ante Jukić¹, Nasser Mohammadiha¹, Toon van Waterschoot², Timo Gerkmann¹, Simon Doclo¹

¹ University of Oldenburg, Dept. of Medical Physics and Acoustics, Oldenburg, Germany

² KU Leuven, Dept. of Electrical Engineering Leuven, Belgium

---

RIRs

R_500: REVERB Challenge, 500 ms

R_700: REVERB Challenge, 700 ms

M_450: MARDY, 450 ms

O_750: Oldenburg, 750 ms

 

Parameters

Sampling frequency f_s = 16 kHz

STFT: tight frame based on 64 ms Hamming window, 16 ms shift

MCLP: length of the filters L_s = 25, number of iterations it_max=5

NMF: IS divergence, rank R=40, multiplicative updates, it_max=100, it_tol=10^-2

---

M=2 microphones

RIRs	Clean	Mic 1	MCLP	MCLP + NMF
R_500
R_700
M_450
O_750

M=4 microphones

RIRs	Clean	Mic 1	MCLP	MCLP + NMF
R_500
R_700
M_450
O_750