The Perl scripts that were used in David Gelbart's PhD thesis to implement hill-climbing feature selection for the Numbers corpus can be found here. (The linked file is in the Unix .tar format, which can be unpacked on Windows using the free 7-Zip tool.) These scripts build on a multi-layer perceptron based Numbers speech recognition system, and (optionally) a noisy version of the Numbers corpus, both of which are available here.
Diagrams of the initial and final feature vectors for hill-climbing in the thesis are here. Plain text lists of the feature vectors can be downloaded here.
After the thesis, we did a technical report that tried hill-climbing with an ensemble of five classifiers instead of three. It can be found here or here. Diagrams of the initial and final feature vectors for hill-climbing with five classifiers are here, and plain text lists of the feature vectors can be downloaded here.
We wrote an INTERSPEECH 2009 conference paper summarizing key parts of the thesis and the technical report.