In this paper we report a study for confidence measure estimation in a large vocabulary speaker-independent continuous speech recognition system. A hybrid confidence measure estimation algorithm was developed. The final confidence measure consists of a number of confidence parameters which are generated from the different processing levels of the recognition system. A Parameter Reliability Analysis (PRA) algorithm was proposed to combine the confidence parameters to form the final confidence measure. The approach was applied to a large vocabulary speaker-independent continuous speech recognition system and obtained superior performance.