A target-independent solution to mandarin speech recognition engine of dialogue system on specified domain

This paper represents a common target-independent solution to mandarin speech recognition engine of the dialogue system on a specified domain. The dialogue system on a specified domain often has many grammar limitations and semantic limitations, and these limitations can be used to guide the process of speech recognition. This paper presents a common solution to the problem of limitation information. Every dialogue system on a specified domain has its own keywords and all these keywords can be clustered into several classes according to their meaning and their parts of speech. Then we can collect sentences on the specified domain to form a sentence corpus and abstract grammar from sentences in this corpus by using these keyword classes. All the grammar can form a finite state network and this network can be used in the decoder process. When this solution is applied to a new domain, what we should do is to replace the keyword definition file and sentences corpus file. By using this solution we can get a detection rate of higher than 85%. When this speech recognition engine was applied to a dialogue system "EasyFlight", a flight information system, a good performance was achieved.