Speaker Recognition Weight Vector Re-estimation via SVM
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing speaker recognition technologies have low identifiability, which affects the accuracy of recognition results.
Innovation Solution
The method involves obtaining score functions using Gaussian Mixture Models (GMMs) and re-estimating a weight vector through a Support Vector Machine (SVM) to improve identifiability by combining likelihood probabilities and posterior probabilities, enhancing the accuracy of recognition results.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional GMM-UBM score function is used for speaker recognition, then the system is simple to implement, but the identifiability is low
Solution Approach 1:
The patent combines GMM-UBM likelihood ratio scoring with SVM classification by merging the score function output with SVM decision function. The score function outputs are used as input features for SVM, creating a hybrid system that leverages both probabilistic modeling and discriminative classification to improve identifiability
Solution Approach 2:
The patent introduces an intermediate score function that transforms GMM-UBM likelihood ratios into a format suitable for SVM processing. This score function acts as a mediator between the probabilistic GMM-UBM model and the discriminative SVM classifier, enabling seamless integration and improving overall recognition accuracy
2Measurement precision
If GMM-UBM likelihood ratio scoring is used, then the recognition process is straightforward, but the recognition accuracy is insufficient
Solution Approach 1:
The patent implements feedback by using SVM to re-estimate the weight vector based on training data. The SVM learns from training examples and adjusts the weighting of different score function outputs, creating a feedback mechanism that continuously optimizes recognition accuracy and improves reliability
Solution Approach 2:
The patent changes the parameter representation by transforming the traditional likelihood ratio scores into a new scoring framework weighted by SVM-learned coefficients. This parameter transformation allows the system to adapt to different speakers and conditions, improving both accuracy and reliability
Data Source
AI summary
A method and device for speaker recognition are provided. In the present invention, identifiability re-estimation is performed on a first vector (namely, a weight vector) in a score function by adopting a support vector machine (SVM), so that a recognition result of a characteristic parameter of a test voice is more accurate, thereby improving identifiability of speaker recognition.