关键词:
摘要: This paper revisits an existing method of constructing high-dimensional probability density functions (PDFs) based on the PDF at output a dimension-reducing feature transformation. We show how to modify so that it can provide with highest entropy among all PDFs generate given low-dimensional PDF. The is completely general and applies arbitrary transformations. chain-rule described for multi-stage calculations typically used in signal processing. Examples are including MFCC auto-regressive features. Experimental verification results using simulated data provided comparison competing generative methods.