Distributed Speech Presence Probability Estimator in Fully Connected Wireless Acoustic Sensor Networks
- PDF / 2,015,909 Bytes
- 21 Pages / 439.37 x 666.142 pts Page_size
- 83 Downloads / 153 Views
Distributed Speech Presence Probability Estimator in Fully Connected Wireless Acoustic Sensor Networks Raziyeh Ranjbaryan1 · Hamid Reza Abutalebi1 Received: 31 May 2019 / Revised: 11 May 2020 / Accepted: 11 May 2020 © Springer Science+Business Media, LLC, part of Springer Nature 2020
Abstract This paper presents a Gaussian-based distributed speech presence probability (DSPP) estimator which is applied in fully connected wireless acoustic sensor networks (WASNs). In WASNs, we are primarily interested in optimally utilizing all available information of recorded signals. In this work, under the Gaussian statistical assumption of signals, each node computes the DSPP using its own local signals along with the compressed signals from other nodes. We evaluate the effect of DSPP estimation on noise reduction from both the simulated and the real recorded signals. The performance of the proposed DSPP estimator is compared to that of local SPP estimation, where each node only uses its noisy signals, and to that of centralized SPP estimation, where each node uses all recorded noisy signals of the whole network. It is shown that the proposed method exhibits good performance, while the computational complexity is considerably reduced. Keywords Speech presence probability · Wireless acoustic sensor networks · Distributed noise reduction algorithms · Gaussian statistical properties
1 Introduction Accurate estimation of speech presence probability (SPP) is required in many speechrelated applications [4,5,8,14,19,22,28,30,33]. In general, a significant improvement in noise reduction from speech is achieved by considering the speech presence uncertainty [4,5,8,22,28,30]. In [8], in each time-frequency unit (TFU), using the SPP, a combination of single-channel Wiener and multi-frame minimum variance distortion-less response filter was proposed, leading to speech quality improvement.
B
Hamid Reza Abutalebi [email protected] Raziyeh Ranjbaryan [email protected]
1
Electrical Engineering Department, Yazd University, Yazd, Iran
Circuits, Systems, and Signal Processing
By incorporating the conditional SPP in the speech distortion weighted multi-channel Wiener filter in [30], an adaptive parameter was introduced, providing a good trade-off between signal distortion and noise reduction. In addition, the accurate estimation of SPP is important in the update of the noise correlation matrix [4,14,19,33]. In [4], a time-varying frequency-dependent parameter based on the SPP was considered, introducing a recursive smoothing technique for noise power spectral density (PSD) estimation. It was shown that the proposed method is able to obtain low estimation error in the case of nonstationary and low signalto-noise ratio (SNR) environments. A soft SPP-dependent minimum mean square error (MMSE)-based technique was proposed in [14], which presents an appropriate estimation of the noise PSD in the case of nonstationary environments. In [19], it was shown that by applying a two-dimensional hidden Markov model and incorporating the spec
Data Loading...
