Abstract

The Delft University of Technology and ING group Netherlands have joint in an effort te create speech recognition tools on a parallel platform. The development of these tools has only just started. As a first step in achieving this goal some speech recognition techniques which have proved themselves, are being ported to a parallel platform, the nCUBE2. One of these techniques is the use of Hidden Markov Models (HMM), a probabilistic method for speech recognition. In this graduation thesis the feasibility of a parallel implementation of the HMM algorithm is studied and the implementation of a first prototype is discussed. It is shown that in the acoustic preprocessing phase as well as in the classification phase of the speech recognizer significant speedups can be reached. In the preprocessing phase especially the Vector Quantization can be done in parallel very efficiently. The training phase of the HMM algorithm can be significantly speeded up. In the recognition phase of the HMM algorithm processor efficiency can reach as high as 78% using 64 processors.

Some speech recognition experiments were carried out with this parallel version of HMM. For single speaker and two speaker data sets good results have been achieved. For multi-speaker environments improvements can be made.