Recognition of Dialog Speech<P>
-- Processing of Spontaneous Speech and Speaker Adaptation --
<P>
Seiichi NAKAGAWA,<P>Mikio YAMAMOTO, Yuji MORIYA, Atsuhiko KAI, Satoshi
KOBAYASHI
<P>
Department of Information and Computer Sciences, Toyohashi University of Technology<P>
Tempaku-cho, Toyohashi, Aichi, 441, Japan<P>
e-mail: nakagawa@say1gw.tutics.tut.ac.jp
<P>
     We investigate a recognition method of spontaneous speech
and an unsupervised speaker adaptation method, and describe
results of analysis of interjections and repairs in real dialog.
     First, we propose and evaluate a processing method of
interjections and unknown words so that a speech recognition
system can deal with spontaneous speech in dialog.  We have
assessed the relationship between the word recognition accuracy
of a system and the detection rate of unknown words by the
simulation of unknown word processing.  We have also evaluated
the performance of our speech recognition system.

     Second, preparing the data to develop a method of processing
of interjections and repairs, we inspected interjections and
repairs in real spoken Japanese dialogs.  Results of analysis
of interfections show that some interjections are used very
frequently and interjections can appear at the boundaries of
Bunsetu(phrase).  Results of analysis of repairs show that
repeating the same words appears to be more common than any other
types of repairs. This type constitutes a little over 30\% of all
repair types. Other types contain substitutions and insertions.

     Third, improving accuracy of speech recognition,
we propose and evaluate an unsupervised speaker adaptation method
on the sequential concatenation training that uses the theory of
MAPE (Maximum A Posteriori probability Estimation) for continuous
parameter HMM.  The label sequences were provided automatically
by the recognizer which worked on a speaker-independent model.
The experimental results show that the better initial model
gives a comparable performance to that of supervised adaptation.
<P>
Keywords: spontaneous speech recognition, unknown word, interjection, repair,
	  unsupervised speaker adaptation<P>