Experimental results for the text-closed data

次へ: Experimental results for the 上へ: Experiments for Continuous Speech 戻る: Experimental conditions for continuous

Experimental results for the text-closed data

The sentence recognition rate was evaluated for the text-closed experiments; the results are shown in Figure 3 . For comparison purposes, results for both the word bigram models and the word trigram models are also shown. In these experiments, the trigram and bigram values were smoothed by deleted-interpolation technique[8].

**図 3:** Recognition rate for text-closed data
$\begin{figure}\begin{center} \fbox{\epsfig{file=figure3.ps,width=70mm}}\end{center}\end{figure}$

In Figure 3 , the vertical axis represents the sentence recognition rate while the horizontal axis represents the number of states. It shows the recognition rate for the 128-state Ergodic HMM is to be higher than that for the word bigram models. It also shows the recognition rate for the 256-state Ergodic HMM to be equal to that for the word trigram models. It is possible though that if the number of states were increased, the recognition rate would have exceeded that for the word trigram models.

**図 4:** Recognition rate for text-open data
$\begin{figure}\begin{center} \fbox{\epsfig{file=figure4.ps,width=70mm}}\end{center}\end{figure}$

次へ: Experimental results for the 上へ: Experiments for Continuous Speech 戻る: Experimental conditions for continuous

Jin'ichi Murakami 平成13年1月19日