Certain telecommunications services [1] currently handled by human representatives are very attractive for automation using automatic speech recognition (ASR). The interactions associated with those services tend to produce sentences with words and phrases (keywords) embedded in continuous speech input. In this paper, we report the development of speaker-independent continuous word recognition (spotting) capabilities to support the automation of such transactions over the public telephone network. The proposed recognition system uses statistical hidden Markov word models for both keywords and non-keywords. The recognition process includes a Viterbi search which segments the input sentence and generates keyword hypotheses followed by a post-decoder that processes the hypothesis segments, computes various a-posteriori likelihood measures, and determines the final outcomes (keyword name, accept / reject).