VAE SPEAKER RECOGNITION FEATURE EXTRACTION METHOD AND SYSTEM USING VARIATIONAL AUTO ENCODER

摘要

The present invention relates to a speaker recognition feature extraction method using a variational auto encoder (VAE) which comprises the following steps of: learning a universal background model (UBM) which is a speaker independent model; generating one integrated vector based on extracted Baum-Welch statistics; learning the VAE with the one integrated vector as an input vector; and extracting latent variables as speaker recognition features. Therefore, the speaker recognition feature extraction method using the VAE can increase speaker recognition performance.