The AT&T Speech-To-Text System is a large vocabulary, continuous speech recognition system with high-accuracy performance and a flexible, modular architecture. We will describe the acoustic, lexical and grammatical modeling and the overall system architecture and search strategy used in this system. In the 1994 ARPA North American Business (NAB) News Evaluation, the system used a 60,000 word vocabulary and thirty-four million 1-5 grams and achieved a word error rate of 10% (H1-P0). We also report results on the ATIS task using the identical system and acoustic models.