This paper gives an overview of the principles of a system for phoneme-based, large-vocabulary, continuous-speech recognition. In particular, the issues of modeling and search for acoustic recognition are addressed. Like many other systems, the recognition architecture is based on an integrated statistical approach. However, the characteristic features of the system are closer to classical pattern recognition: 1. Classical pattern recognition techniques like continuous mixture densities and linear discriminant analysis are extensively used, resulting in what can be viewed as a sort of 'statistical template matching'. 2. The framework of Hidden Markov modeling is used for time alignment only. 3. Time-synchronous beam search is used consistently throughout all tasks; by using a tree organization of the vocabulary and phoneme look-ahead, this one-pass strategy is able to handle a 10000-word task.