The Deterministic Boltzmann Machine (DBM) is a form of neural network that learns much faster than the original stochastic Boltzmann Machine that it is derived from. In this paper we overview briefly the theory of DBMs, and describe their application to speech recognition. In a static phoneme configuration task the DBM obtained an average recognition rate of 98.6 % (best: 99.1 %) for the "bdg" task, and 97 % for an all-consonant task. In a dynamic recognition task (including time-shifts), rates are less good by a few percent, but a state-feedback dynamic architecture provided some improvement.