Identity Vector Processing for Speaker Recognition Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Identity vectors used in speaker recognition are prone to interference from channel variability and environment variability, leading to reduced accuracy in speaker identity verification.
Innovation Solution
A method and device that process identity vectors by selecting interclass and intraclass neighboring vectors, determining differences, and using basis vectors to maximize interclass difference while minimizing intraclass difference, thereby improving recognition accuracy through feature transformation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If identity vectors are used directly for speaker recognition, then the recognition process is simple, but the accuracy is reduced due to interference from channel and environment variability
Solution Approach 1:
The patent segments the identity vector processing into multiple distinct components: obtaining identity vectors from multiple speakers, selecting interclass neighboring identity vectors, selecting intraclass neighboring identity vectors, determining interclass difference, determining intraclass difference, and determining basis vectors. This segmentation transforms a single complex recognition task into manageable sub-tasks that collectively improve accuracy while controlling overall complexity.
Solution Approach 2:
The patent introduces a new dimensional approach by selecting neighboring identity vectors from different classes (interclass) and same class (intraclass), then determining differences in these additional dimensions. This dimensional expansion allows the system to capture more discriminative features beyond the original identity vectors, improving recognition accuracy by considering relationships in multiple vector spaces.
2Measurement precision
If interclass and intraclass neighboring identity vectors are selected and processed, then recognition accuracy is improved, but the processing complexity increases
Solution Approach 1:
The patent performs preliminary actions by pre-selecting interclass and intraclass neighboring identity vectors before the actual recognition decision. The basis vectors are determined in advance through the difference analysis, allowing the system to prepare discriminative features beforehand. This preliminary processing organizes the complex computations into structured steps that can be efficiently executed and reused.
Solution Approach 2:
The patent extracts only the essential discriminative information by selecting specific neighboring identity vectors (interclass and intraclass) and determining their differences. Instead of processing all available identity vectors, the system extracts and utilizes only the most relevant neighboring vectors that contribute to distinguishing between different speakers, thereby reducing unnecessary computational overhead.
Data Source
AI summary
Processing circuitry of an information processing apparatus obtains a set of identity vectors that are calculated according to voice samples from speakers. The identity vectors are classified into speaker classes respectively corresponding to the speakers. The processing circuitry selects, from the identity vectors, first subsets of interclass neighboring identity vectors respectively corresponding to the identity vectors and second subsets of intraclass neighboring identity vectors respectively corresponding to the identity vectors. The processing circuitry determines an interclass difference based on the first subsets of interclass neighboring identity vectors and the corresponding identity vectors; and determines an intraclass difference based on the second subsets of intraclass neighboring identify vectors and the corresponding identity vectors. Further, the processing circuitry determines 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.