70 likes | 238 Views
Bayesian Networks Toolkit. Objective: free C++ toolkit to manipulate and experiment with Bayesian Networks Status: Released at the INRIA Gforge: BaNeTo A simple C++ interface has been designed to access the underlying functionalities
E N D
Bayesian Networks Toolkit • Objective: free C++ toolkit to manipulate and experiment with Bayesian Networks • Status: • Released at the INRIA Gforge: BaNeTo • A simple C++ interface has been designed to access the underlying functionalities • But: Continuous training in some cases with hidden variables still buggy ! • Issue: No more engineer to fix this bug…
Missing Data Recognition • Achievements during the last period: • Design and implementation of a new (theoretically more correct) marginalisation procedure • Design and implementation of a new masks model with temporal and frequency constraints • Validation of the proposed approach on the Aurora2 database
Missing Data Recognition:New marginalisation • Local SNR < 0dB => data maskedLocal SNR > 0dB => data unmasked • This translates into:with
Missing Data Recognition:New masks model • Masks model = HMM • Observation pdfs (GMMs) = p(y|m,state) • Transition probabilities = Frequency constraints • Frequency constraints: • Clustering of the space covered by training masks • Observation pdf: m = one of this VQ codebook • First time the correlation between the feature dimensions are modeled !