A high-performance, flexible, and potentially inexpensive speech recognition system is described. The system is based on two special-purpose integrated circuits that perform the speech recognition algorithms very efficiently. One of these integrated circuits is the front-end processor, which computes spectral coefficients from incoming speech. The second integrated circuit computes a dynamic-time-warp algorithm. The system can compare an input word with 1000-word templates and respond to a user within s. The system demonstrates that computational complexity need not be a major limiting factor in the design of speech recognition systems.