Comparison of the R&B and DSM approaches
HBR 14/07/1998


The underlying models in the R&B and DSM approaches
Fig 1. The R&B model
Fig 2. The DSM model
Segmentation and labelling:  Phone segments.
Targets:  Targets in learnt hidden space. 
Filter:  Arbitrary fixed dynamics.
Hidden representation:  Learnt hidden space. 
Non-linear mapping:  Multi-layer perceptron (MLP). 
Surface acoustic form:  Mel-cepstra (MFCC).
Sources of variability:  Gaussian noise in observations.
 
Segmentation and labelling:  Adjusted phone segments 
Targets:  Targets in formant hidden space 
Filter:  Dynamics learnt via EKF 
Hidden representation:  Formants, but adapts. 
Non-linear mapping:  Multi-layer perceptron (MLP) 
Surface acoustic form:  Mel-cepstra (MFCC) 
Sources of variability:  Gauss. noise in observns and targets.