When I start to read people's source code and papers, it usually gives me hints to solve the same problem using different ways. Advances in speech recognition research caused many people forget that a speech recognizer's performance is closely related to its implementation. Many people forget that flaws in implementation will still cause significant degradation in performance.
My advice is we should feel thankful to any open-source speech recognizer in the world because all of them allows us to have better idea what should we do in the future.
If you are interested what I think about them, follow here .
Hidden Markov Model Toolkit (HTK)
ISIP Foundation Class and Speech Recognition System (IFC)
OpenCV . Check out the audio-visual CSR code.