#include <srch_flat_fwd_internal.h>
Public Attributes | |
char const * | word_cand_dir |
char const * | latfile_ext |
int32 | word_cand_win |
word_cand_t ** | word_cand |
int32 | n_word_cand |
hmm_context_t * | hmmctx |
whmm_t ** | whmm |
word_ugprob_t ** | word_ugprob |
backoff_t * | ug_backoff |
backoff_t * | filler_backoff |
uint8 * | tg_trans_done |
int32 * | rcscore |
s3wid_t * | word_cand_cf |
ctxt_table_t * | ctxt |
fwd_dbg_t * | fwdDBG |
latticehist_t * | lathist |
int32 | n_frm |
int32 | final_state |
int32 | renormalized |
int32 | multiplex |
int32 | multiplex_singleph |
pctr_t * | ctr_mpx_whmm |
pctr_t * | ctr_nonmpx_whmm |
pctr_t * | ctr_latentry |
ptmr_t | tm_hmmeval |
ptmr_t | tm_hmmtrans |
ptmr_t | tm_wdtrans |
kbcore_t * | kbcore |
Referenced by whmm_eval().
Referenced by whmm_eval().
A context table. This parameter, as well as the one in word_fsg.c should all go up to srch.c. They are more generic then the others.
Referenced by flat_fwd_dag_add_fudge_edges(), whmm_exit(), whmm_transition(), word_enter(), and word_trans().
Filler probability
Referenced by word_trans().
Final state is supposed to be the last state, so it is usually equal to n_state-1
Debug object of srch_flat_fwd.c
Structures for flat lexicon decoding search The HMM context.
Referenced by whmm_eval(), whmm_transition(), and word_enter().
A pointer for convenience
Referenced by dump_all_whmm(), dump_all_word(), flat_fwd_dag_add_fudge_edges(), whmm_eval(), whmm_exit(), whmm_renorm(), whmm_transition(), word_enter(), and word_trans().
char const* srch_FLAT_FWD_graph_s::latfile_ext |
Complete word candidate filename for an utterance formed by word_cand_dir/<uttid>.latfile_ext
Viterbi history (backpointer) table
Whether we will use multiplexed triphones
Whether we will use multiplexed triphones
Number of frame of this utternance
Referenced by dump_all_word(), whmm_eval(), whmm_exit(), whmm_transition(), word_enter(), and word_trans().
#candidate entries in word_cand for current utterance. If <= 0; full search performed for current utterance
Referenced by word_trans().
rc scores uncompacted; one entry/rc-ciphone
Referenced by word_trans().
Whether scores had to be renormalized in current utt
Referenced by whmm_renorm().
If tg_trans_done[w] TRUE, trigram transition to w occurred for a given history, and backoff bigram transition from same history should be avoided
Unigram backoff probability
Referenced by word_trans().
The word hmms list. For actual search traverse
Referenced by whmm_eval(), and word_enter().
Word candidates for each frame. (NOTE!! Another array with a hard limit on its size.)
Referenced by word_trans().
BAD_S3WID terminated array of candidate words for word transition in current frame (if using input word lattices to restrict search).
Referenced by word_trans().
char const* srch_FLAT_FWD_graph_s::word_cand_dir |
Structures for decoding utterances subject to given input word lattices; ie, restricting the decoding to words found in the lattice. (For speeding up the decoding process.) NOTE: This mode is optional. If no input lattice is given, the entire vocabulary is eligible during recognition. Also, SILENCEWORD, FINISHWORD, and noisewords are always eligible candidates.
Input lattice specifies candidate words that may start at a given frame. In addition, this forward pass can also consider words starting at a number of neighbouring frames within a given window.
Input lattice file format: Each line contains a single <word> <startframe> info. The line may contain other info following these two fields; these are ignored. Empty lines and lines beginning with a # char in the first column (ie, comment lines) are ignored. Directory containing candidate words files. If NULL, full search performed for entire run
In frame f, candidate words in input lattice from frames [(f - word_cand_win) .. (f + word_cand_win)] will be the actual candidates to be started(entered)
Referenced by word_trans().
word unigram probability
Referenced by word_trans().