00010		The Stanford B-4 Phonetic Transcription System
00015	
00100	The Stanford B-4 Speech Recognition System is a preprocessor that
00200	attempts to extract the maximum amount of useful information from an
00300	acoustic input stream. It does this by making use of accumulated
00400	information obtained during training sessions.  A' priori
00500	probabilities based on these data are computed and stored in
00600	signature tables, which correlate parameters extracted from the
00700	acoustic stream with the phonemic content of the utterance.
00800	
00900	Three types of tables are used, one type to provide for speaker
01000	normalization, a second type to relate the normalized input
01100	parameters to the phonetic intent of the speaker and the third type
01200	to transcribe strings of phonetic events into sequences of phonemes.
01250	Several output sequences are supplied, with probability ratings for the
01300	complete strings as well as for the individual phonemes.
01400	
01500	Heavy use is made of the ability of the signature tables to express
01600	complicated multi-modal probabilistic rlationships between sets
01650	of input parameters and the intended outputs, and to formulate these
01800	relationships automatically from examples supplied in training runs.
02000	
02100	The purpose of the B-4 is not to provide a complete speech
02200	understanding system but rather to perform some very useful
02300	preprocessing, using very fast techniques, before the acousticly
02400	derived clues are combined with the supplementary semantic, syntactic
02500	and contextual information that may be available to the complete
02600	system.