We describe here a novel 'language processor' (LP) which uses syntactic and semantic properties of language for optimising the performance of a Hindi (an Indian language) Speech recognition system for railway reservation enquiry task. The acoustic level recognizer provides several alternatives for each word with associated 'confidence levels'. Using these, the LP has to generate a 'most likely' sentence consistent with the constraints imposed by syntax and semantics. The LP requires no apriori knowledge of the syntax or semantics of the language and acquires this during a training phase. During the recognition phase, it uses this knowledge at the sentence level to correct for word recognition errors made by the acoustic level recognizer.