#21079First, we present a new paradigm for speaker-independent (SI) training of hidden Markov models (HMM), which uses a large amount of speech from a fewspeakers instead of the traditional practice of using a little speech from many speakers.
other,10-8-H90-1060,ak
is transformed to the space of the
<term>
target speaker
</term>
and combined by averaging . Using
#21239Each reference model is transformed to the space of thetarget speaker and combined by averaging.
tech,33-3-H90-1060,ak
many
<term>
speakers
</term>
prior to
<term>
training
</term>
. With only 12
<term>
training speakers
#21127In addition, combination of the training speakers is done by averaging the statistics of independently trained models rather than the usual pooling of all the speech data from many speakers prior totraining.
lr,20-4-H90-1060,ak
word error rate
</term>
on a standard
<term>
grammar
</term>
and
<term>
test set
</term>
from the
<term>
#21149With only 12 training speakers for SI recognition, we achieved a 7.5% word error rate on a standardgrammar and test set from the DARPA Resource Management corpus.
model,1-7-H90-1060,ak
<term>
new ( target ) speaker
</term>
. A
<term>
probabilistic spectral mapping
</term>
is estimated independently for each
#21211Aprobabilistic spectral mapping is estimated independently for each training (reference) speaker and the target speaker.
other,22-4-H90-1060,ak
a standard
<term>
grammar
</term>
and
<term>
test set
</term>
from the
<term>
DARPA Resource Management
#21151With only 12 training speakers for SI recognition, we achieved a 7.5% word error rate on a standard grammar andtest set from the DARPA Resource Management corpus.
lr-prod,26-4-H90-1060,ak
</term>
and
<term>
test set
</term>
from the
<term>
DARPA Resource Management corpus
</term>
. This performance is comparable
#21155With only 12 training speakers for SI recognition, we achieved a 7.5% word error rate on a standard grammar and test set from theDARPA Resource Management corpus.
other,3-9-H90-1060,ak
combined by averaging . Using only 40
<term>
utterances
</term>
from the
<term>
target speaker
</term>
#21249Using only 40utterances from the target speaker for adaptation, the error rate dropped to 4.1% --- a 45% reduction in error compared to the SI result.
other,16-7-H90-1060,ak
reference ) speaker
</term>
and the
<term>
target speaker
</term>
. Each
<term>
reference model
</term>
#21226A probabilistic spectral mapping is estimated independently for each training (reference) speaker and thetarget speaker.
tech,9-9-H90-1060,ak
the
<term>
target speaker
</term>
for
<term>
adaptation
</term>
, the
<term>
error rate
</term>
dropped
#21255Using only 40 utterances from the target speaker foradaptation, the error rate dropped to 4.1% --- a 45% reduction in error compared to the SI result.
lr,27-2-H90-1060,ak
</term>
, which uses a large amount of
<term>
speech
</term>
from a few
<term>
speakers
</term>
instead
#21075First, we present a new paradigm for speaker-independent (SI) training of hidden Markov models (HMM), which uses a large amount ofspeech from a few speakers instead of the traditional practice of using a little speech from many speakers.
lr,23-6-H90-1060,ak
corpus
</term>
and a small amount of
<term>
speech
</term>
from the
<term>
new ( target ) speaker
#21201Second, we show a significant improvement for speaker adaptation (SA) using the new SI corpus and a small amount ofspeech from the new (target) speaker.
other,15-5-H90-1060,ak
<term>
test suite
</term>
, using 109
<term>
training speakers
</term>
. Second , we show a significant
#21175This performance is comparable to our best condition for this test suite, using 109training speakers.
model,14-2-H90-1060,ak
speaker-independent ( SI ) training
</term>
of
<term>
hidden Markov models ( HMM )
</term>
, which uses a large amount of
<term>
#21062First, we present a new paradigm for speaker-independent (SI) training ofhidden Markov models ( HMM ), which uses a large amount of speech from a few speakers instead of the traditional practice of using a little speech from many speakers.
tech,7-1-H90-1060,ak
paper reports on two contributions to
<term>
large vocabulary continuous speech recognition
</term>
. First , we present a new paradigm
#21042This paper reports on two contributions tolarge vocabulary continuous speech recognition.
other,26-6-H90-1060,ak
amount of
<term>
speech
</term>
from the
<term>
new ( target ) speaker
</term>
. A
<term>
probabilistic spectral mapping
#21204Second, we show a significant improvement for speaker adaptation (SA) using the new SI corpus and a small amount of speech from thenew ( target ) speaker.
other,9-7-H90-1060,ak
is estimated independently for each
<term>
training ( reference ) speaker
</term>
and the
<term>
target speaker
</term>
#21219A probabilistic spectral mapping is estimated independently for eachtraining ( reference ) speaker and the target speaker.
other,10-5-H90-1060,ak
comparable to our best condition for this
<term>
test suite
</term>
, using 109
<term>
training speakers
#21170This performance is comparable to our best condition for thistest suite, using 109 training speakers.
measure(ment),12-9-H90-1060,ak
</term>
for
<term>
adaptation
</term>
, the
<term>
error rate
</term>
dropped to 4.1 % --- a 45 % reduction
#21258Using only 40 utterances from the target speaker for adaptation, theerror rate dropped to 4.1% --- a 45% reduction in error compared to the SI result.
tech,22-3-H90-1060,ak
models
</term>
rather than the usual
<term>
pooling
</term>
of all the
<term>
speech data
</term>
#21116In addition, combination of the training speakers is done by averaging the statistics of independently trained models rather than the usualpooling of all the speech data from many speakers prior to training.