Research Topic

Robust Distributed Multi-Source Detection and Labeling in Wireless Acoustic Sensor Networks

An example of a WASN consisting of 7 speech sources (red) and 20 nodes (blue).

The ceaseless increase in complexity of signal processing tasks urge involving a modern fashion that considers solving multiple heterogeneous nodes that cooperate in various signal processing tasks in a wireless sensor network. This notion is quite disparate from the ongoing Information and Communication Technologies canvas, where the lack of superior performance advances the intent of a compelling MDMT paradigm.

In the particular studied audio case, multiple sensors at each device record speech primarily from a specific speaker of interest. This will arise many challenges mainly the existence of impulsive noise, mixed signals and the labeling of source task in order to keep nodes well informed about the interest of their neighborhood.

Aim

VAD results (red lines) for the estimated speech signal S7 using our proposed approach "Distributed stability selection based sparseness and robust Mahalanobis classifier for VAD" (DSRM-VAD) with a correct detection rate of 98.9%.

Performing a robust distributed Voice Activity Detection for the described scenario is necessary to surpass dealing with cooperative nodes for a unique signal processing task in the classical approaches . Moreover, the appeal of holding up the performance level of the designed detector when abandoning some assumptions that are not sustained in practice.

Our aim is to develop robust distributed multi-source detection techniques in scenarios where nodes have node-specific interests. This requires detecting sources of interest for the different nodes in adverse environments taking into account robust measures. Distributed multi-source detection is a new research field and an important enabler for MDMT systems.

Furthermore, we target developing robust distributed feature extraction which can be used for multi-source detection. The goal is to robustify the selected features and make them applicable in adverse MDMT scenarios. Precisely, our techniques should be robust with respect to uncertainties in the noise distribution and adaptive, to cope with environment non-stationarities.

 

 

Key Research Area

Multi-source detection; multi-source labeling; source separation; energy unmixing; multiple devices for multiple tasks; wireless acoustic sensor networks; classification; voice activity detection; robust feature extraction; distributed algorithms; sparse coding; non-negative independent component analysis.

Contact

Lala Khadidja Hamaidi
M.Sc.

Address:

Dolivostra├če 15

D-64293 Darmstadt

Germany

Phone:

+49 6151 16 - 24385

Fax:

+49 6151 16 - 24404

Office:

S4|10-206

Email:

hamaidi (at) gsc.tu...

 Print |  Impressum |  Sitemap |  Search |  Contact |  Privacy Policy
zum Seitenanfangzum Seitenanfang