Language Model Selection based on the Analysis of Japanese Spontaneous Speech on Travel Arrangement Task

This paper deals with the issue of language model selection based on the analysis of data collection for spontaneous speech in Japanese in the travel rrangement task which contains five different subtasks. The procedure of transcription and segmentation of the Japanese spontaneous speech in Romanized transcription is described. The use of topic-dependent separated language model were evaluated in calculating the perplexity and applying it into Japanese speech recognition of the travel arrangement task corpus. The reduction of perplexity was shown and the increase of speech recognition was performed by use of the subtopic language model