Experimental conditions

We developed a word HMM to connect phone HMMs. To train continuous mixture HMMs, 2,635 word utterances were used. For the training of bigram probabilities, 8,475 sentences of the ATR Dialog Database were used in text-open experiments and smoothed by the deleted-interpolation algorithm [4]. Then the same 8,475 sentences plus 38 test sentences were used in text-closed experiments. In this way, it was possible to perform both the text-open and text-closed experiments using the same test data. To reduce the memory requirements, we used beam pruning. The results of these experiments were evaluated in terms of the word correct rate and word accuracy rate[3]. The experimental conditions are summarized in Table 1.

表 1: Experimental conditions

phone model	3-state 4-loop
	continuous mixture HMM
mixture number	max 10 ( valid for each syllable )
acoustic parameter	16th order LPC cepstrum + power
	+$\Delta$ power + 16th order $\Delta$ cepstrum
frame window	20 ms
frame period	5 ms
training voice	word speech (2,620 words)
# syllable categories	26 syllables
language model	word bigram
training data	8,475 sentences (57,354 words)
vocabulary	435
beam width	4,096
test sentence count	38 sentences