Analysis and effect of speaking style for dialogue speech recognition

This paper analyzes acoustic likelihood calculated from two acoustic models, a spontaneous speech acoustic model and a read speech acoustic model, from the viewpoint of linguistic information, such as word category and language likelihood. Experimental results show a significant tendency in the relationship between speaking style and linguistic information. According to the analysis results, a word's acoustic likelihood calculated from the spontaneous speech acoustic model is higher, or more suitable, than that from the read speech acoustic model in the case when the word is an interjection or an auxiliary verb. On the other hand, even in human-to-human conversation, a word's acoustic likelihood calculated from the read speech acoustic model can be higher than that from the spontaneous speech acoustic model in the case when the word is a noun. Applying this knowledge along with machine learning, post-processing experiments of the results of ASR using these two acoustic models are carried out. In this set of experiments, post-processing, based on a support vector machine, is applied. The experimental results show that the selection scheme, based on word category, reduces word error rate by 1.62 points over the single system.