文档介绍：Ecien t Algorithms for Sp eec h Recognition
Mosur K. Ra vishank ar
Ma y 15, 1996
CMU-CS-96-143
Sc ho ol puter Science
Computer Science Division
Carnegie Mellon Univ ersit y
Pittsburgh, P A 15213
Submitte dinp artial ful l lment of the r e quir ements
for the de gr e eofDo ctor of Philosophy.
mittee:
Rob erto Bisiani, co-c hair Univ ersit y of Milan
Ra j Reddy , co-c hair
Alexander Rudnic ky
Ric hard Stern
W a yne W ard
c
1996 Mosur K. Ra vishank ar
This researc hw as supp orted b y the Departmen t of the Na vy ,Na v al Researc h Lab oratory under
Gran t No. N00014-93-1-2005. The views and conclusions con tained in this do cumen t are those of
the author and should not b e in terpreted as represen ting the ocial p olicies, either expressed or
implied, of the . go v ernmen t.
Keyw ords: Sp eec h recognition, searc h algorithms, real time recognition, lexical
tree searc h, lattice searc h, fast matc h algorithms, memory size reduction.
Abstract
Adv ances in sp eec h tec hnology puting p o w er ha v e created a surge of
in terest in the practical application of sp eec h recognition. Ho w ev er, the most accurate
sp eec h recognition systems in the researc hw orld are still far to o slo w and exp ensiv eto
b e used in practical, large v o cabulary con tin uous sp eec h applications. Their main goal
has b een recognition accuracy , with emphasis on acoustic and language mo delling.
But practical sp eec h recognition also requires putation to b e carried out in
real time within the limited resources|CPU p o w er and memory size|monly
a v puters. There has b een relativ ely little w ork in this direction while
preserving the accuracy of researc h systems.
In this thesis, w e fo cus on ecient and ac cur ate sp eec h recognition. It is easy to
impro v e recognition sp eed and reduce memory requiremen ts b y trading a w a y accu-
racy , for example b y greater pruning, and using simpler acoustic and lan