NOTE: Yes, there is CMU-Sphinx.
This project provides basic "speech" recognition system (digits only) starting from preparing studying sample for Neural Network from given database of sound to software that records audio from system input and tries to categorize it using trained Neural Network.
Project was written as diploma software for my customer.
Neural Network library used is Encog.
Code contains parts of SpectroEdit project and funf-open-sensing-framework.
First user must prepare database of samples of sounds with following directory structure:
<Dictor>/<Number>/<Samplefile#1>.wav <Dictor>/<Number>/<Samplefile#2>.wav
It is important not to mess these files up, as this will degrade recognition performance greatly.
Main application then uses it to build a database of features, using selected FeatureExtractor. This produces two output files: one with FeatureExtractor parameters and other with .csv file suitable for use in Encog to train a Multi-Layer Perceptron.
Most sucessfull FeatureExtractor is MFCC -- Mel-Frequency Cepstrum Coefficients.
Second step is to run recognizer (either from NetBeans or with appropriate command). Recognizer requires FeatureExtractor paremeters' file as well as file with trained neuron network. GUI is semi-obviously.