The work described in this paper addresses the problem for extracting bispectrum feature of speech data. Very often the bispectrum feature extraction and data reduction are complicated due to some limiting constraints, i.e., no prior knowledge of feature's distribution and higher dimensionality of bispectrum data. In this article we developed an adaptive feature extraction mechanism based on cascade neural network in conjunction with feature's dimensionality reduction based on Karhunen-Loeve transformation technique. An adaptive codebook generation algorithm which is a cascade configuration of SOFM (Self Organizing Feature Map) and LVQ (Learning Vector Quantization) was used before the K-L transformation. The transformation was experimentally shown as an effective procedure for orthogonalization and dimensionality reduction of bispectrum feature. Performance of our speaker identification system was perceived to be significantly increased eventhough using limited number of channels in noisy environment. We also tried to improve the capability of adaptive codebook generation algorithm by applying simplified differential competitive learning (SDCL) network.
|Number of pages||7|
|Journal||Proceedings of SPIE - The International Society for Optical Engineering|
|Publication status||Published - 2000|
|Event||Hybrid Image and Signal Processing VII - Orlando, FL, USA|
Duration: 25 Apr 2000 → 25 Apr 2000