E�cien t Searc h Strategies in Hierarc hical P attern Recogniti on Systems by Neera j Deshm ukh Joseph Picone Y u-Hung Kao Dept. of EE Inst. for Signal & Info. Pro c. Systems & Info. Sci. Lab. Boston Univ ersit y Mississippi State Univ ersit y T exas Instrumen ts Inc. Boston, MA 02215 MS State, MS 39762 Dallas, T exas 75243 ne er aj@engc.bu.e du pic one@e e.msstate.e du yhkao@csc.ti.c om SSST, '95 Marc h 13, 1995
The Problem of Statistical P attern Recognition � Mathematical represen tation d p ( W = A ) = max p ( W = A ) W Ba y es' Theorem: d p ( W = A ) = arg max p ( A=W ) p ( W ) W � Automatic Sp eec h Recognition � Acoustic Mo del � Language Mo del � Searc h � Hidden Mark o v Mo dels � Complex Applications ) Hierarc hical Mo deling Large V o cabulary Con tin uous Sp eec h Recognition (L V CSR) � T arget recognition in radars, SAR imagery � In telligen t database access (audio-video) �
Sp eec h Recognition System Speech Input Speech Input Digital Signal Processing Model Structure Digital Signal Processing Test Data Acoustic Training Data Language Models Models Learning Algorithm (Restimation of parameters) Hypothesis Generation the cat in the hat bat in the mat Trained Models SEARCH Training of Models in Speech Recognition Final recognized word sequence " the cat in the hat got the rat." Continuous Speech Recognition System Figure 1: T raining and Recognition Schematic of training and recognition systems
Searc h Strategies � Searc h P aradigm T o c ho ose a pattern with the highest lik eliho o d score for our feature mo dels giv en the observ ed data. � Motiv ation The n um b er of h yp otheses (c hoices for the correct pattern) gro ws exp onen tially with n um b er of feature mo dels. Hence a strategy that sa v es on computation and storage requiremen ts is sough t. � P opular searc h tec hniques Viterbi Searc h � Viterbi Beam Searc h � ? � A Stac k Deco ding � N-b est Searc h > Main tains h yp otheses within sp eci�ed b eam. al l > Propagates top N h yp otheses at eac h state. > N is indep enden t of Viterbi b eam. > T o ol to in tegrate information from m ultiple sources. > P artial to w ards shorter h yp otheses. Generalized N-b est Searc h � > F orw ard-Bac kw ard Searc h > Progressiv e Lattice Searc h
Application to Hierarc hical Systems Sentence level HMM structure Word level Phone level Hierarchical Model for Speech Recognition Figure 2: Hierarc hical structure of feature mo dels � Multi-lev el computation. Information �o w b oth across and along la y ers of mo del � framew ork. � Excessiv e requiremen ts on computation and storage capacit y . F or N-b est, N di�eren t paths are to b e traced bac k � ) degradation of p erformance of Viterbi searc h. � In tegration of di�eren t N-b est h yp otheses obtained from di�eren t lev els.
F rame-Sync hronous Viterbi Searc h � Motiv ation Reduce computational and memory requiremen ts b y limiting the n um b er of v alid h yp otheses to b e pro cessed. � FSVS Algorithm Nodes having scores above FSVS pruning threshold Nodes pruned at this frame, as score < FSVS pruning threshold nodes in order of path score frames Frame Synchronous Viterbi Pruning Figure 3: FSVS pruning algorithm During Viterbi b eam searc h, at the end of eac h frame of input data 1. Sort all activ e h yp otheses in decreasing order of scores. 2. Keep only a few top-scoring h yp otheses based on some pruning threshold. This threshold can b e �xed or dynamic , dep ending on the application.
Practical Issues � Computational o v erhead � The p erils of o v er-pruning � Goal: Generalized N-b est Algorithm � should iden tify mo dels asso ciated with output sym b ols. should k eep N highest scoring paths for suc h mo dels at � ev ery lev el in the mo del hierarc h y . should trace all these paths to obtain N b est h yp otheses � at the top lev el. The generalized N-b est searc h for a hierarc hical system � th us main tains N-b est h yp otheses at every lev el and not only at the top. � Adv an tages � Reduction in problem space � V ery useful in memory-critic al applications � Some gain in computational e�ciency
Exp erimen tal Results 1200 1200 1000 delta_slew = 2*max_delta 1000 delta_slew = 2*max_delta max_delta = 0.3 max_delta = 0.3 prune_hyp_factor = 1.5 800 800 prune_hyp_thresh = 35 active sds active sds 600 600 400 400 200 200 0 0 0 200 400 600 800 1000 0 200 400 600 800 1000 frames frames a: Viterbi b eam sea rch b: FSVS Figure 4: Memory usage for (sp eec h) sen tence recognition T yp e Sent. Comptn. Mem. W o rd Sent. of over (fracn of slots/ erro r erro r p runing �o w realtime) frame % % Viterbi 308 0.289 590 24.2 36.1 FSVS 0 0.274 424 3.5 21.2 Results of F rame-Sync hronous Viterbi Searc h
Summary Hierarc hical pattern recognition systems are required to � solv e complex pattern matc hing applications. � Suc h systems ha v e excessiv e requiremen ts of memory and computational p o w er for searc h. F rame-sync hronous Viterbi Searc h algorithm is a step in � reducing the problem space. � FSVS algorithm is particularly attractiv e to memory- critical recognition tasks.
Recommend
More recommend
