Experimental conditions for sentence speech recognition

For training of a continuous mixture HMM, 2635 word utterances were used. The number of states for the Ergodic HMM as a language model was set to 2, 4, and 8. In addition, the sentence likelihood is the product of the acoustic model likelihood and the with the power of the language model likelihood $n$, in our experiments we used $n = 32$.

We used 38 sentences as test data. For training the Ergodic HMM, in the text-open experiment we used 4000 sentences of the ATR Dialog Database, whereas in the text-closed experiment we used the same 4000 sentences plus the 38 test sentences. In this way, we can perform the text-closed experiment and a text-open experiment in the same test data.

The experimental conditions are shown in more detail in Table 3 .

表 3: Experimental conditions for sentence speech recognition

#syllable models	52
Syllable model	4-state 3-loop Gaussian mixture
	continuous HMM
Learning data	male announcer, 2635 word utterances
Parameter	log power + 16 order LPC-cepstrum
	+ $\Delta$log power + 16 order $\Delta$ cepstrum
Test data	same speaker, 38 sentences
Vocabulary	435
Beam width	1024
Speech style	read speech
#state of	2,4,8
Ergodic HMM
learning data	4000 sentences, 57354 words