There are two aspects to this question, really. First, the deep one, is: what is the process behind developing new features? How can we make the design decisions? How can we optimize the parameters? This is more than I want to take on here, but see Brian Kingsbury's Ph.D. thesis for one way to pursue this process.
The second question: Now I that I have this great idea for new features, how can I calculate them for a standard corpus so I can try them on a real recognition task? Often this happens through modification of the standard feature calculation programs such as feacalc or rasta. However, for a radically new feature calculation, you can take advantage of the neat framework developed by Brian, as described in the drspeech_featools man page.
Previous: 5.6 What are delta features? How do you calculate them? - Next: 5.8 How do different features compare performancewise?
Back to ICSI Speech FAQ index
Generated by build-faq-index on Tue Mar 24 16:18:16 PDT 2009