46 * Merged from branch: SPHINX3_5_2_RCI_IRII_BRANCH: this fills in the code for doing tree propagation. However, rescoring is still being used. The code is working. However, it takes up huge amount of memory and I consider this as not elegant. It also shows that straight forward implementation of tree copies search doesn't work even in these days.
53 * 1, Added an experimental version of tree expansion interface it the code, it does tree expansion without history pruning. Currently disabled because it used to much memory space srch_word_switch_tree.[ch]. 2, Remove -lminsearch segments of code, it proves to be unnecessary. 3, Remove the rescoring interface. In this search, WST_rescoring is actually not doing rescoring, it is rather a segment of code which collect all active word end together and input it into the viterbi history.
59 * Incorporated several fixes to the search. 1, If a tree is empty, it will be removed and put back to the pool of tree, so number of trees will not be always increasing. 2, In the previous search, the answer is always "STOP P I T G S B U R G H </s>"and filler words never occurred in the search. The reason is very simple, fillers was not properly propagated in the search at all <**exculamation**> This version fixed this problem. The current search will give <sil> P I T T S B U R G H </sil> </s> to me. This I think it looks much better now.
62 * 1, Turn on mode 5 again, 2, fixed srch_WST_end, 3, Add empty function implementations of add_lm and delete_lm in mode 5. This will make srch.c checking happy.
68 * Log. Implementation of word-switching tree. Currently only work for a
69 * very small test case and it's deliberately fend-off from users. Detail
70 * omitted.
71 *
72 * Revision 1.9 2005/05/11 06:10:39 archan
73 * Code for lattice and back track pointer table dumping is now wrapped in reg_result_dump. The function is shared across mode 4 and mode 5. Possibly later for mode 3 and mode 6 as well.
74 *
75 * Revision 1.8 2005/05/03 04:09:10 archan
76 * Implemented the heart of word copy search. For every ci-phone, every word end, a tree will be allocated to preserve its pathscore. This is different from 3.5 or below, only the best score for a particular ci-phone, regardless of the word-ends will be preserved at every frame. The graph propagation will not collect unused word tree at this point. srch_WST_propagate_wd_lv2 is also as the most stupid in the century. But well, after all, everything needs a start. I will then really get the results from the search and see how it looks.
77 *
78 * 17-Mar-2005 A. Chan (archan@cs.cmu.edu) at Carnegie Mellon University
79 * Started. Word condition tree search. Aka lexical tree copies.