Pause adaptation

Even though a pause skip technique is employed, the insertion error in the pause period did not always disappear. So, we retrained the $/pause/$ HMM for the test data with the Baum-Welch algorithm using some $/pause/$ units at the beginning of the test data. Because of this procedure, the sentence recognition rate was 83.9% as opposed to 61.7% with no such procedure for the SI case. This recognition rate was very close to the SD case's. A summary of these experiments is shown in Table 3 .

sentence recognition rate(%)

表 3: Results of read speech recognition

language		bigram	bigram	trigram	trigram
procedure	order	SD	SI	SD	SI
original	1	42.5%	0.0%	66.7%	0.0%
word trigram	$\sim$8	51.3%	0.0%	75.1%	0.0%
pause skip	1	49.4%	31.4%	71.6%	61.7%
	$\sim$8	60.2%	44.4%	80.0%	76.7%
pause skip &	1	60.5%	44.8%	90.4%	83.9%
adaptation	$\sim$8	76.2%	55.6%	97.7%	96.6%

text-closed; beam width: 4096; $\alpha$: 1;