I am a fresher in speech recognition. I want to apply VTLN to
telephone speech corpus.
according to HTKBook, at first, I set the configuration variable
WARPFREQ 0.8 , run the tools hcopy to get mfcc, and run herest to get
a hmm model;
Then change WARPFREQ 0.85, run hcopy again to get another mfcc, run
herest with single pass re-estimate above the first HMM model
but i find the output of herest no change :
- average log prob per frame = -1.088801e+002
- total frames seen = 1.392573e+006
i don't know how to do next?
is it wrong with my experiments about VTLN in HTK?
WARPFREQ should be estimate individually or only one WARPFREQ value
exist globally ?
Any suggestion will be appreciated. Thanks