[PDF] - Ecien t Searc h Strategies in Hierarc hical P attern PDF Document

SLIDE 1 Ecien t Searc h Strategies in Hierarc hical P attern Recogniti

n

Systems by Neera j Deshm ukh Joseph Picone Y u-Hung Kao Dept.

f

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

ne@e

e.msstate.e du yhkao@csc.ti.c

m

SSST, '95 Marc h 13, 1995

SLIDE 2 The Problem

f

Statistical P attern Recognition

Mathematical

represen tation p( d W = A) = max W p(W = A) Ba y es' Theorem: p( d W = A) = arg max W p(A=W )p(W )

Automatic

Sp eec h Recognition

Acoustic

Mo del

Language

Mo del

Searc

h

Hidden

Mark

v

Mo dels

Complex

Applications ) Hierarc hical Mo deling

Large

V

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)

SLIDE 3 Sp eec h Recognition System

Digital Signal Processing Speech Input Learning Algorithm (Restimation of parameters) Digital Signal Processing Trained Models Model Structure Training Data

Training of Models in Speech Recognition

Speech Input Language Models Acoustic Models Hypothesis Generation the cat in the hat bat in the mat SEARCH Final recognized word sequence the cat in the hat got the rat." "

Continuous Speech Recognition System

Test Data

Figure 1: T raining and Recognition Schematic

f

training and recognition systems

SLIDE 4 Searc h Strategies

Searc

h P aradigm T

c

ho

se

a pattern with the highest lik eliho

d

score for

ur

feature mo dels giv en the

bserv

ed data.

Motiv

ation The n um b er

f

h yp

theses

(c hoices for the correct pattern) gro ws exp

nen

tially with n um b er

f

feature mo dels. Hence a strategy that sa v es

n

computation and storage requiremen ts is sough t.

P
pular

searc h tec hniques

Viterbi

Searc h

Viterbi

Beam Searc h

A

? Stac k Deco ding

N-b

est Searc h > Main tains al l h yp

theses

within sp ecied b eam. > Propagates top N h yp

theses

at eac h state. > N is indep enden t

f

Viterbi b eam. > T

l

to in tegrate information from m ultiple sources. > P artial to w ards shorter h yp

theses.
Generalized

N-b est Searc h > F

rw

ard-Bac kw ard Searc h > Progressiv e Lattice Searc h

SLIDE 5 Application to Hierarc hical Systems

Sentence level Word level Phone level

Hierarchical Model for Speech Recognition

HMM structure

Figure 2: Hierarc hical structure

f

feature mo dels

Multi-lev

el computation.

Information
w

b

th

across and along la y ers

f

mo del framew

rk.
Excessiv

e requiremen ts

n

computation and storage capacit y .

F
r

N-b est, N dieren t paths are to b e traced bac k ) degradation

f

p erformance

f

Viterbi searc h.

In

tegration

f

dieren t N-b est h yp

theses
btained

from dieren t lev els.

SLIDE 6 F rame-Sync hronous Viterbi Searc h

Motiv

ation Reduce computational and memory requiremen ts b y limiting the n um b er

f

v alid h yp

theses

to b e pro cessed.

FSVS

Algorithm

Nodes having scores above FSVS pruning threshold Nodes pruned at this frame, as score < FSVS pruning threshold frames nodes in order

f path score

Frame Synchronous Viterbi Pruning

Figure 3: FSVS pruning algorithm During Viterbi b eam searc h, at the end

f

eac h frame

f

input data 1. Sort all activ e h yp

theses

in decreasing

rder
f

scores. 2. Keep

nly

a few top-scoring h yp

theses

based

n

some pruning threshold. This threshold can b e xed

r

dynamic, dep ending

n

the application.

SLIDE 7 Practical Issues

Computational
v

erhead

The

p erils

f
v

er-pruning

Goal:

Generalized N-b est Algorithm

should

iden tify mo dels asso ciated with

utput

sym b

ls.
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

btain

N b est h yp

theses

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

theses

at every lev el and not

nly

at the top.

Adv

an tages

Reduction

in problem space

V

ery useful in memory-critic al applications

Some

gain in computational eciency

SLIDE 8 Exp erimen tal Results

200 400 600 800 1000 200 400 600 800 1000 1200 frames active sds delta_slew = 2*max_delta max_delta = 0.3 200 400 600 800 1000 200 400 600 800 1000 1200 frames active sds delta_slew = 2*max_delta max_delta = 0.3 prune_hyp_factor = 1.5 prune_hyp_thresh = 35

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

rd

Sent.

f
ver

(fracn

f

slots/ erro r erro r p runing

w

realtime) frame % % Viterbi 308 0.289 590 24.2 36.1 FSVS 0.274 424 3.5 21.2 Results

f

F rame-Sync hronous Viterbi Searc h

SLIDE 9 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

f

memory and computational p

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.