Voice Characteristic Extraction via Joint Loss Optimization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current systems for determining voice characteristics, such as speaker verification and identification, face challenges in maximizing between-class differences and minimizing intra-class variations, leading to suboptimal performance in classification and clustering tasks.
Innovation Solution
The implementation of a computer-implemented method that jointly minimizes a non-sampling-based loss function for between-class separability and a Gaussian mixture loss function with a non-unit multi-variant covariance matrix for intra-class variability reduction, using a deep learning model with convolutional and residual network layers to optimize voice characteristic extraction.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional machine learning algorithms are used for speaker classification, then the system can perform basic voice recognition, but the between-class differences are not maximized and intra-class differences are not minimized, leading to suboptimal classification performance
Solution Approach 1:
The patent transforms the classification problem by changing the parameter space from raw acoustic features to learned embedding representations. By training the model to output embeddings that maximize between-class differences and minimize intra-class differences, the system achieves better classification accuracy and generalization capability simultaneously
Solution Approach 2:
The patent introduces an embedding layer as an intermediary between the acoustic feature extraction and the classification process. This embedding layer learns to map acoustic features to a representation space where classification is more effective, acting as a mediator that transforms the data to resolve the contradiction between accuracy and generalization
2Measurement precision
If deep learning models with complex architectures are used to improve voice characteristic extraction, then the model can capture more detailed acoustic representations, but the computational complexity and training difficulty increase
Solution Approach 1:
The patent segments the deep learning model into distinct functional modules: acoustic feature extraction layers, embedding layers, and loss function components. This segmentation allows each module to be optimized independently and simplifies the overall training process by breaking down the complex architecture into manageable parts
Solution Approach 2:
The patent designs a universal model architecture that can perform multiple functions: acoustic feature extraction, embedding learning, speaker classification, and speaker clustering. By making the model multi-functional, the system achieves high extraction accuracy without needing separate specialized models for each task, thereby reducing overall complexity
3Measurement precision
If the model is trained to optimize speaker classification, then between-class separability improves, but intra-class variation is not adequately reduced, limiting clustering performance
Solution Approach 1:
The patent merges the classification objective and clustering objective into a single unified training framework. By combining the between-class separability optimization (classification loss) and intra-class consistency optimization (clustering loss) into a joint training process, the model simultaneously achieves both goals without compromising either
Solution Approach 2:
The patent changes the training parameters by introducing a composite loss function that incorporates both classification and clustering objectives. This parameter change allows the optimization process to simultaneously maximize between-class differences and minimize intra-class variations, resolving the contradiction between the two objectives
4Ease of manufacture
If Gaussian mixture models with unit covariance matrices are used for speaker clustering, then the clustering process is simplified, but the model cannot accommodate diverse speaker embeddings distributions, reducing clustering accuracy
Solution Approach 1:
The patent makes the covariance matrix dynamic by allowing it to be learned from data rather than fixed to a unit matrix. This dynamic approach enables the model to adapt the covariance structure to the actual distribution of speaker embeddings, improving clustering accuracy while maintaining reasonable implementation complexity through the use of diagonal constraints
Data Source
AI summary
Methods, systems, and apparatus, including computer programs encoded on computer storage media, for determining voice characteristics are provided. One of the methods includes: obtaining speech data of a speaker; inputting the speech data into a model trained at least by jointly minimizing a first loss function and a second loss function, wherein the first loss function comprises a non-sampling-based loss function and the second loss function comprises a Gaussian mixture loss function with non-unit multi-variant covariance matrix; and obtaining from the trained model one or more voice characteristics of the speaker.