This paper presents a keyword spotting system based on the NSHP-HMM. This model allows to dynamically create global word models from letters models, and do not require any writing segmentation. The second section describes our system and its application to a keyword-based handwritten mail sorting task. Next section shows how to divide process- ing time by 4, using a fix-point arithmetic and a dynamic model desactivation approach based on the natural length complexity. First results are encouraging, particulary for a document-level analysis.