Disclosed are a voice recognition system and a construction method for the voice recognition system. By way of layering the system, a general semantic recognition operation for the system is separated from a specific semantic recognition operation for an application program; and by way of classifying the application programs and abstracting out a common performance function, the system can find the application program matching the voice content semantics very efficiently and a third-party program is easily added into the existing voice recognition system. The present invention maps the performance function to a regular expression with semantic variables, so that the system can recognize more semantic expression manners with optimization of the semantic recognition. Therefore, the system can show more humanized characteristics.