Voiceprint Recognition Model Using Composite Loss Functions
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current speaker recognition systems based on convolutional neural networks (CNNs) using softmax loss functions often suffer from overfitting, resulting in good performance on training sets but poor performance on untrained test sets.
Innovation Solution
The implementation of a voiceprint recognition method that jointly optimizes a voiceprint recognition model using a normalized exponential function and a centralization function to reduce intra-class variation and improve discrimination between speakers, thereby enhancing recognition performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a softmax loss function is used for training a CNN-based speaker recognition system, then the system achieves good performance on the training set, but the system suffers from overfitting and shows poor performance on untrained test sets
Solution Approach 1:
The patent changes the loss function parameters from a single softmax loss to a composite loss function combining softmax loss and center loss. The center loss introduces a new parameter (center point for each class) that pulls features of the same speaker closer together, reducing intra-class variation and preventing overfitting while maintaining training accuracy.
Solution Approach 2:
The patent creates a composite loss function by combining two different loss functions (softmax loss and center loss) with weighted summation. This composite approach leverages the strengths of both functions: softmax loss for inter-class discrimination and center loss for intra-class compactness, thereby improving generalization performance.
2Ease of manufacture
If traditional speaker recognition systems are designed to maintain simplicity, then the system structure remains easy to implement, but the system lacks the capability to effectively reduce intra-class variation and improve speaker discrimination
Solution Approach 1:
The patent modifies the training objective parameters by introducing center loss with simple center point calculations for each speaker class. This parameter change enables the system to reduce intra-class variation without requiring complex architectural modifications, maintaining implementation simplicity while improving discrimination capability.
Solution Approach 2:
The patent segments the loss function into two distinct components (softmax loss for between-speaker discrimination and center loss for within-speaker compactness). This segmentation allows each component to address specific aspects of speaker recognition independently, improving overall performance while keeping the implementation modular and manageable.
Data Source
AI summary
Embodiments of this application disclose a voiceprint recognition method performed by a computer. After obtaining a to-be-recognized target voice message, the computer obtains target feature information of the target voice message by using a voice recognition model, the voice recognition model being obtained through training according to a first loss function and a second loss function. Next, the computer determines a voiceprint recognition result according to the target feature information and registration feature information, the registration feature information being obtained from a voice message of a to-be-recognized object using the voiceprint recognition model. The normalized exponential function and the centralization function are used for jointly optimizing the voice recognition model, and can reduce an intra-class variation between depth features from the same speaker. The two functions are used for simultaneously supervising and learning the voice recognition model, and enable the depth feature to have better discrimination, thereby improving recognition performance.