Speaker Embedding Network for Voice Uncertainty Robustness
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing speaker recognition systems lack robustness to noisy data and voice uncertainty, as they do not adequately consider feature uncertainty and distribution in voice data, leading to low performance in real-world conditions.
Innovation Solution
A learning device and method that utilizes a speaker embedding extracting neural network with sequentially coupled networks to calculate features, accuracies, and speaker embeddings, considering voice uncertainty by modeling hidden variables and uncertainties using Gaussian distributions, thereby enhancing robustness.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a general x-vector is used for speaker embedding extraction, then the system structure is simple, but the robustness to random perturbation and voice uncertainty is low
Solution Approach 1:
The neural network is divided into multiple distinct networks (first network for feature calculation, second network for accuracy calculation, third network for posterior distribution calculation, fourth network for speaker embedding calculation) that process different aspects of the input signal. This segmentation allows each network to specialize in handling specific uncertainties while maintaining overall system manageability.
Solution Approach 2:
The patent introduces a new dimension of uncertainty estimation by calculating not only speaker features but also accuracy metrics and posterior distributions. This adds probabilistic information as an additional dimension to the traditional deterministic speaker embedding approach, enabling the system to account for voice uncertainty without fundamentally redesigning the entire architecture.
2Measurement precision
If weighting is performed to assign large weight to important feature vectors, then the classification accuracy improves, but the robustness to voice data under actual conditions remains low due to not considering uncertainty
Solution Approach 1:
The patent changes the parameters being optimized from only speaker features to include both speaker features and accuracy metrics. By modifying the loss function to consider uncertainty parameters alongside traditional classification targets, the system achieves both high classification accuracy and robustness to noisy data through unified optimization.
Solution Approach 2:
The accuracy calculation network provides feedback about the reliability of extracted features back to the overall learning process. This feedback mechanism allows the system to adjust its predictions based on uncertainty estimates, improving robustness while maintaining accuracy through iterative optimization of the joint loss function.
Data Source
AI summary
Learning means 81 generates a speaker embedding extracting neural network by learning a weighting factor so as to minimize a loss function indicating an error between a speaker label indicating a speaker of a voice signal and an output value output from an output layer, with respect to a neural network including an input layer that receives an input of the voice signal and the output layer that outputs the output value indicating a speaker of the voice signal. The speaker embedding extracting neural network includes a network that calculates a first accuracy from a feature in units of frames and calculates an average and a second accuracy of a posterior distribution from an average and an accuracy of a prior distribution and the feature and the first accuracy.