Title :
On the effect of topological structure of the Kohonen network on the performance of a hierarchical two layered isolated word recognition system
Author :
Kepuska, V.Z. ; Gowdy, John N.
Author_Institution :
Dept. of Electr. & Comput. Eng., Clemson Univ., SC
Abstract :
A novel hierarchical neural network system for isolated word recognition is presented. Currently, it can simultaneously perform recognition of different acoustic segments, i.e., recognize both phonemes and isolated words, and use them to attain more accurate operation. Simulation results show that one of the significant factors affecting the performance of the system is the topological structure of the self-organizing feature map units. This fact brings up the issues related to data representation in interfacing the layers of the hierarchical system. Experiments demonstrate that data representation of this interface must preserve the topological relationship of self-organizing feature map units. This system is able to achieve a 100% recognition score for the task of a small but difficult 20-isolated-word vocabulary with very limited training. Five utterances of each word are used for training and the remaining five for testing of a single speaker
Keywords :
hierarchical systems; knowledge representation; learning systems; neural nets; speech recognition; topology; Kohonen network; acoustic segments; data representation; hierarchical system; isolated word recognition; learning systems; neural network; phonemes; self-organizing feature map; speech recognition; topological structure; Automatic testing; Band pass filters; Computational modeling; Filter bank; Hierarchical systems; Neural networks; Speech recognition; System testing; Vector quantization; Vocabulary;
Conference_Titel :
Southeastcon '90. Proceedings., IEEE
Conference_Location :
New Orleans, LA
DOI :
10.1109/SECON.1990.117771
