Identity vector processing method and computer device

A computer device, comprising:

• processing circuitry configured to:
• obtain, a set of identity vectors that are calculated according to voice samples from speakers, the identity vectors being classified into speaker classes respectively corresponding to the speakers;
• select, from the identity vectors, first subsets of interclass neighboring identity vectors respectively corresponding to the identity vectors, a first subset of the first subsets of interclass neighboring identity vectors corresponding to an identity vector of the identity vectors being in different speaker classes from the corresponding identity vector;
• select, from the identity vectors, second subsets of intraclass neighboring identity vectors respectively corresponding to the identity vectors, a second subset of the second subsets of intraclass neighboring identity vectors corresponding to an identity vector of the identity vectors being in a same speaker class as the corresponding identity vector;
• determine an interclass difference based on the first subsets of interclass neighboring identity vectors and the corresponding identity vectors;
• determine an intraclass difference based on the second subsets of intraclass neighboring identity vectors and the corresponding identity vectors;
• determine a set of basis vectors to maximize a projection of the interclass difference on the basis vectors and to minimize a projection of the intraclass difference on the basis vectors;
• convert an input identity vector of a target speaker according to the basis vectors;
• and
• perform a verification of the target speaker based on the converted input identity vector.

A non-transitory computer-readable medium storing instructions which when executed by a computer cause the computer to perform:

• obtaining a set of identity vectors that are calculated according to voice samples from speakers, the identity vectors being classified into speaker classes respectively corresponding to the speakers;
• selecting, from the identity vectors, first subsets of interclass neighboring identity vectors respectively corresponding to the identity vectors, a first subset of the first subsets of interclass neighboring identity vectors corresponding to an identity vector of the identity vectors being in different speaker classes from the corresponding identity vector;
• selecting, from the identity vectors, second subsets of intraclass neighboring identity vectors respectively corresponding to the identity vectors, a second subset of the second subsets of intraclass neighboring identity vector corresponding to an identity vector of the identity vectors being in a same speaker class as the corresponding identity vector;
• determining an interclass difference based on the first subsets of interclass neighboring identity vectors and the corresponding identity vectors;
• determining an intraclass difference based on the second subsets of intraclass neighboring identity vectors and the corresponding identity vectors;
• determining a set of basis vectors to maximize a projection of the interclass difference on the basis vectors and to minimize a projection of the intraclass difference on the basis vectors;
• converting an input identity vector of a target speaker according to the basis vectors;
• and
• performing a verification of the target speaker based on the converted input identity vector.