This paper deals with the problem of detection keywords/rejection out-of-vocabulary in continuous speech recognition. Two different techniques based on confidence measures are investigated to improve the detection of keywords using a grammar founded on loop of phones. Confidence measures are computed from phone level information provided by a Hidden Markov model based speech recognizer. We use two kinds of likelihood as ratio and distance, and theirs normalised forms to compute a confidence measures for each word. All confidence measures are are evaluated using the French SPEECHDAT database. The Figure-Of-Merite (FOM) for the normalized likelihood ratio is about $68.2\%$ compared to $71.5\%$ obtained by the normalized likelihood distance