Model-based distributed node clustering and multi-speaker speech presence probability estimation in wireless acoustic...

Area of Science:

• Signal Processing
• Acoustic Sensor Networks
• Machine Learning

Background:

• Speech presence probability (SPP) is crucial for noise estimation and speech enhancement.
• Existing SPP estimation methods are primarily single-channel or centralized multi-channel.
• SPP estimation in wireless acoustic sensor networks (WASNs), especially with multiple speakers, presents significant challenges.

Purpose of the Study:

• To develop a distributed model-based SPP estimation method for multi-speaker detection in WASNs.
• To eliminate the need for a central fusion center in SPP estimation.
• To enhance speech detection robustness in challenging acoustic environments.

Main Methods:

• A distributed k-means clustering algorithm is employed to group sensor nodes into subnetworks for individual speaker detection.
• Local estimation of speech and noise power spectral densities is performed at each node within subnetworks.
• A distributed consensus method facilitates both distributed clustering and SPP estimation.

Main Results:

• The proposed distributed clustering effectively assigns nodes to subnetworks based on noisy observations.
• The distributed SPP estimator demonstrates robust speech detection capabilities across various noise conditions.
• The method successfully addresses multi-speaker scenarios in WASNs without centralized processing.

Conclusions:

• The developed distributed SPP estimation method is effective for multi-speaker detection in WASNs.
• The approach offers a decentralized solution, overcoming limitations of centralized systems.
• This work advances SPP estimation techniques for complex acoustic sensor network applications.