Based on the demisyllableas the recognition primitive(unit), in this paper, a 3-stage recognition structure is presented. It combines the pattern matching method based on speech production model with the feature rule method based on"knowledge", and permit