Experimental condition

To test an Ergodic HMM as an SNG, experiments were performed. As training datum, we used the ATR Dialog Database. The example sentences are shown in Table  1 . The learning conditions of the experiments are shown in Table 2 .

表 1: Example of sentences

$\bullet$ hai moshimoshi

(Hello)

$\bullet$ e-tto sochira dai ichi kai no tsuuyaku deNwa kokusai kaigi no

jimu kyoku de sho u ka

(Well, is this the office of the first international conference on

interpreting telephony ?)

$\bullet$ hai sou desu ( Yes )

$\bullet$ e-tto chotto sono kaigi no koto dene

(Well about this conference ...)

$\bullet$ hai douzo (Yes...)

$\bullet$ e-tto ima temoto ni ano touroku youshi ga aru N desu keredomo

(I have a registration form, now.)

表 2: Learning conditions for Ergodic HMM

type of Hidden Markov model	Mealy type
learning sentence number	8475
total word number	57354 words
number of states	2, 4, 8 and 16
vocabulary	6418
Initial state distribution	equivalent value
Initial transition value	random
Initial symbol value	random
end of HMM learning	under 1% likely hood

図 1: Results of an 8-state Ergodic HMM analysis