This may be interesting for some of us. It's Open Source.

Audio-visual continuous speech recognition (AVCSR) uses a set of visual and
audio features to increase the accuracy of speech recognition in noisy
environments. The AVCSR package uses OpenCV and contains a set of Visual C++
functions for visual feature extraction as well as binary code for audio
feature extraction (MFC coefficients) and the fusion model (coupled hidden
Markov model). A set of test .avi files are also provided. To run the
appliaction, download and install avcsr.zip and unzip the oficexxx.avi files
under \test\AvsrDemo\WinData. 

https://sourceforge.net/project/shownotes.php?group_id=22870&release_id=1546
35