Fourth European Conference on Speech Communication and Technology

Madrid, Spain
September 18-21, 1995

Parallel Implementation of an Hybrid Neural Network Used for Speech Recognition Task

T. Calonge (1), L. Alonso (1), R. Ralha (2), A. L. Sanchez (1)

(1) Dpto. de Informatica, Universidad de Valladolid, Facultad de Ciencias, c/ Prado de la Magdalena s/n, Valladolid, Spain
(2) Dpt. Matematica. Campus del Gualtar, Universidade do Minho, Braga, Portugal

In this paper a parallel implementation of a learning process for the two main non recurrent neural networks will be presented: Multilayer Perceptron (MLP) and Self-Organising Map of Kohonen (SOM). They will be integrated in just one system used for isolated word recognition task, speaker independent, with a restricted vocabulary, made up of the ten Spanish words for the decimal digits. The implementation was carried out on a distributed memory multiprocessor system with a synchronous message passing communication network. A detailed explanation about the computational tasks distribution and the subsequent communication requirements will be given. In addition, several experiments with different numbers of processors were made in order to study both the granularity and the overhead due to the processors communication. In this way, we obtained the speedup curve showing the load balance and efficiency.

Full Paper

Bibliographic reference.  Calonge, T. / Alonso, L. / Ralha, R. / Sanchez, A. L. (1995): "Parallel implementation of an hybrid neural network used for speech recognition task", In EUROSPEECH-1995, 153-156.