Negative filter
Lincoln, CSR, system 14
(448.2 per million)
tech,17-3-H91-1010,ak
error rate
</term>
by about 10 % in both
<term>
triphone and semiphone systems
</term>
. A new
<term>
training strategy
</term>
#21757A very simple improved duration model has reduced the error rate by about 10% in bothtriphone and semiphone systems.
tech,27-4-H91-1010,ak
improvements can be obtained by a related
<term>
rapid adaptation technique
</term>
. Finally , the
<term>
recognizer
</term>
#21789A new training strategy has been tested which, by itself, did not provide useful improvements but suggests that improvements can be obtained by a relatedrapid adaptation technique.
model,9-5-H91-1010,ak
recognizer
</term>
has been modified to use
<term>
bigram back-off language models
</term>
. The
<term>
system
</term>
was then
#21802Finally, the recognizer has been modified to usebigram back-off language models.
other,8-7-H91-1010,ak
results
</term>
are presented for both the
<term>
RM and ATIS CSR tasks
</term>
. This paper describes an implemented
#21838Evaluation test results are presented for both theRM and ATIS CSR tasks.
other,11-6-H91-1010,ak
from the
<term>
RM task
</term>
to the
<term>
ATIS CSR task
</term>
and a limited number of
<term>
development
#21818The system was then transferred from the RM task to theATIS CSR task and a limited number of development tests performed.
measure(ment),0-7-H91-1010,ak
development tests
</term>
performed .
<term>
Evaluation test results
</term>
are presented for both the
<term>
RM
#21830The system was then transferred from the RM task to the ATIS CSR task and a limited number of development tests performed.Evaluation test results are presented for both the RM and ATIS CSR tasks.
measure(ment),9-3-H91-1010,ak
duration model
</term>
has reduced the
<term>
error rate
</term>
by about 10 % in both
<term>
triphone
#21749A very simple improved duration model has reduced theerror rate by about 10% in both triphone and semiphone systems.
tech,3-5-H91-1010,ak
adaptation technique
</term>
. Finally , the
<term>
recognizer
</term>
has been modified to use
<term>
bigram
#21796Finally, therecognizer has been modified to use bigram back-off language models.
tech,2-4-H91-1010,ak
and semiphone systems
</term>
. A new
<term>
training strategy
</term>
has been tested which , by itself
#21764A newtraining strategy has been tested which, by itself, did not provide useful improvements but suggests that improvements can be obtained by a related rapid adaptation technique.
model,4-3-H91-1010,ak
been tested . A very simple improved
<term>
duration model
</term>
has reduced the
<term>
error rate
</term>
#21744A very simple improvedduration model has reduced the error rate by about 10% in both triphone and semiphone systems.
tech,4-2-H91-1010,ak
system
</term>
. Some new variations in
<term>
semiphone modeling
</term>
have been tested . A very simple
#21734Some new variations insemiphone modeling have been tested.
tech,19-6-H91-1010,ak
task
</term>
and a limited number of
<term>
development tests
</term>
performed .
<term>
Evaluation test
#21826The system was then transferred from the RM task to the ATIS CSR task and a limited number ofdevelopment tests performed.
other,7-6-H91-1010,ak
</term>
was then transferred from the
<term>
RM task
</term>
to the
<term>
ATIS CSR task
</term>
and
#21814The system was then transferred from theRM task to the ATIS CSR task and a limited number of development tests performed.
tech,1-6-H91-1010,ak
back-off language models
</term>
. The
<term>
system
</term>
was then transferred from the
<term>
#21808Thesystem was then transferred from the RM task to the ATIS CSR task and a limited number of development tests performed.