Experiments for spontaneous speech

We perform experiments with word trigram models and for the SD case. The test data were spoken as follows.

1. Read speech
   Speakers uttered text data in read style.

2. Semi spontaneous speech
   Speakers uttered text data that contained symbols indicating filled-pauses. Except for the filled-pauses, the above utterances were the same as the read speech.

3. Spontaneous speech
   Speakers remembered the intention of the text data, and spoke freely. The contents of this utterances differed from the read speech. Therefore, this test data included filled-pauses and out-of-vocabulary words.

The context of the text data was the same as in section 4.4; the speakers, however, differed and were not broadcast announcer.