next up previous
次へ: Discussion 上へ: Experiments for Continuous Speech 戻る: Experimental results for the

Experimental results for the text-open data

The results of the text-open continuous speech recognition experiments are shown in Figure  4 . For comparison purposes, results for the word trigram models are also shown. In these experiments, both the sentence recognition rate and word accuracy rate[7] were evaluated.

In the figure, a 48.6% sentence recognition rate is obtained for the 128-state Ergodic HMM. On the other hand, the recognition rate for the word trigram models is 47.1%. And the word accuracy rate of the 16-state Ergodic HMM which is equal to that of the word trigram models. These results show that the Ergodic HMM can outperform word trigram models for text-open data. On the other hand, the 256-state Ergodic HMM has lower performance compared to the word trigram models. This means that this model does overturning for text data.



Jin'ichi Murakami 平成13年10月2日