Bruce Lowerre, 1995 This system is shareware. As such, it is supplied for educational and personal use. It is not to be sold or used in any commercial products without my consent. Portions of code in some modules contain code proprietary to Entropic. See ./doc/MGR_HTK_License for restrictions on this code. This is the MGR system. This system includes the following: the MGR system architecture for software development real-time processing of speech signals real_time speech signal end point detection speaker independent recognition connected word recognition phoneme level speech recognition word level speech recognition automatic phoneme labeling of utterance files automatic word labeling of utterance files gender recognition wave form X-window graphics routines spectrogram X-window graphics routines example live input with circular buffer a bnf to network compiler As Eric Thayer would say, "MGR is both a floor polish and a desert toping!" WARNING: There is a LOT here and will require some time to absorb it all. See README.INSTALL for installing the system. The MGR system is a general development system which can be used for developing algorithms for any large complex AI system. This architecture has evolved over 25 years. The particular configuration here is for speech recognition. Thus, there are actually two large topics to learn, the MGR development system and speech recogntion. See the subdirectory called "doc" for the following infomation: running.mgr - infomation on getting started running the system. mgr - this mgr configuration for speech recogntion The algorithms used are suitable for either speaker dependent or indepentent, isolated word, or connected word recognition. The algorithms are designed for "real-time" processing. That means that the input signal is processed in parrallel to being "captured." To do live input, an end point detector is included. An example live input module, sunlive, is included. The sun audio system is NOT recommended; it's the PITS! The sunlive module is included only to serve as an example of how to capture a live input signal within the MGR architecture. The trained HMM models provided in the demos have been trained using the HTK software. With the permission of Entropic (see ./doc/MGR_HTK_License), MGR includes parts of the HTK signal processing routines as well as code to provide compatibility with HTK files. See ./doc/ABOUT_HTK for information on Entropic's HTK software. The particular HMM models can be either phonetic or word and are processed identically by the MGR recognition routines. The difference depends only on the bnf grammar. The parameters used in this system are for 8KHz signals (specifically, telephone bandwidth) and use 10 msec frame steps with 128 sample point windows. 10 Melcep parameters are calculated plus energy, the first order difference, and the second order difference for a total of 33 parameters per frame. This is a single mixture, continuous density HMM (using a diaganol covariance matrix). This gives adequate performance for the demos and shows how the system works which is the intent of this software package. As the primary design goal of MGR is modularity, it is my hope that this software can serve as the basis for many speech research projects and can help overcome the "Tower of Babble" problem of communication among the many research groups. If everyone (speech research groups) all use a common development system, then the exchange of information and algorithms will be greatly facilitated. I hope that other people can contribute MGR modules such as live input for various hardware platformes. Please send them to me and I shall include them in future releases of MGR. See the files in the doc subdirectory for more information.