We have derived an expression for the integral of the DAF-HMM likelihood function with respect to the un-augmented features. This allows normalizing the DAF-HMM likelihood function so that it can be compared with likelihood functions based on the un-augmented features. We demonstrated the use of the method on two data sets. In particular, we have shown that appending feature time derivatives achieves lower classification error as well as higher average log-likelihood for data with slowly-varying spectral character. For data with abrupt spectral character, the opposite was observed. This indicates the possibility of a ``litmus test" for the use of DAF. It suggests the possibility of using DAF-HMM along with HMM togeter in a single classifier.

Baggenstoss 2017-05-19