In this paper we describe a preliminary investigation of the use of fillers at the lexical level rather than the modelling level of a hidden Markov model-based keyword spotter and a detector of new words. In our last system, keywords and out-of-vocabulary speech shared the same context-dependent phoneme models with no explicit modelling of extraneous speech. Thus a task-independent training is performed for all models while the scoring method uses a two-pass Viterbi-type algorithm based on a lexical tree constructed with transcriptions of keywords and fillers using the same set of 40 English phonemes. The distinction between keywords and extraneous speech is performed during the search by using the lexical tree and language models. Thus using a simple method, we perform a faster training and allow easier modifications for the word-spotting task. On the other hand this kind of architecture allows our system to be used for both keyword spotting and new word detection tasks.