Unsupervised Extraction of Meaningful Nonlinear Principal Components Applied for Voice Conversion

摘要

Nonlinear Principal Component Analysis (NLPCA) is one of the most progressive computational tools developed during the last two decades. However, in spite of its proper performance in feature extraction and dimension reduction, it is considered as a blind processor which can not extract physical or meaningful factors from dataset. This paper presents a new distributed model of autoassociative neural network which increases meaningfulness degree of the extracted parameters. The model is implemented to perform Voice Conversion (VC) and, as it will be seen through comparisons, results in proper conversion quality.