This paper describes the design and development of the architecture of dedicated hardware which recognizes continuous speech sentences using a 1000-word vocabulary. It is based on the HMM-LR mechanism, which is an integration system of speech recognition and language analysis. Many real-time techniques using 33 DSPs for multi-processing were developed for the heavy computation of the trellis algorithms and state duration control. The paper gives performance data with results within a pipeline delay of approximately 2 or 3 seconds, independently of the length of the sentence.