Research

We are developing nonlinear estimation approaches which rapidly adapt on-line to speaker variations and changing noise sources. This is accomplished by utilizing only the noisy speech of interest to create nonstationary network models which can be used to remove noise from the given signal.  Methods includes:

• Dual Extended Kalman Filtering.  A dual state-space approach based on nonlinear extensions to Kalman filter theory. The method involves the concurrent estimation of both the underlying clean speech and the parameters of the network model. Includes general research on the prediction and estimation of noisy time series.
  
  
• Unscented Kalman Filtering.  A fundemental algorithmic advance over the Extended Kalman Filter.  Includes general research on state-estimation, paramater estimation, and dual-estimation.
  
  
• Noise Regualrized Adaptive Filtering.  A neural filter which directly maps from noisy speech to enhanced speech, and utilizes a new regularizer that constrains the network weights and allows for direct training with only the noisy speech signal.
  
  

Review chapter on-line: Networks For Speech Enhancement (ps version)

Active noise cancellation is an approach based the principle of destrcutive inteference to perform the physical cancellation and control of sound and vibration.

• Adjoint LMS.  A computationally efficient method for multi-input-multi-output active noisecancellation systems.