Internet-accessible speech recognition technology
Kun Huang, Joseph Picone · 2003
Speech recognition systems can be viewed as an application of complex pattern recognition and machine learning algorithms. The development of such a system is a time-consuming and infrastructure-intensive task. The Institute for Signal and Information Processing (ISIP) developed one of the first fully-functional public domain speech recognition systems. In this paper, we introduce the major components of this system which include: a digital signal processing front end that generates feature vectors from the speech signal, a hidden Markov Model (HMM) trainer which estimates acoustic model parameters and a hierarchical search decoder which implements an efficient time-synchronous Viterbi beam search.