Speech Presence Probability Estimation for Small-Array MVDR Headphones
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing speech enhancement technologies using the MVDR algorithm suffer from low accuracy in determining speech presence probability, especially with small microphone arrays, leading to poor performance in devices like headphones with limited microphone spacing.
Innovation Solution
A method and system that optimize speech presence and absence models using maximum likelihood estimation (MLE) and expectation maximization (EM) algorithms to improve the accuracy of speech presence probability estimation, enabling better convergence and noise covariance matrix estimation for enhanced speech enhancement.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a voice activity detection (VAD) algorithm is used to estimate speech presence probability, then the noise covariance matrix can be determined, but the estimation accuracy is relatively low, especially when the count of microphones is small
Solution Approach 1:
The patent changes the parameters of the probability models by introducing entropy as an optimization criterion. The entropy of the speech presence probability and speech absence probability are calculated and used to guide the optimization process, transforming the estimation approach from direct VAD-based methods to an entropy-maximization framework that achieves higher accuracy with small microphone arrays.
Solution Approach 2:
The patent replaces the traditional mechanical/statistical VAD algorithm with an information-theoretic approach based on entropy calculation. By substituting the conventional probability estimation mechanism with an entropy-based optimization system, the method achieves superior accuracy without requiring additional hardware complexity.
2Reliability
If the MVDR algorithm is used with small microphone arrays, then the device complexity is reduced, but the speech enhancement effect drops sharply
Solution Approach 1:
The patent optimizes the probability distribution parameters by introducing entropy maximization as the optimization criterion. This parameter transformation allows the system to achieve reliable speech enhancement with fewer microphones by extracting more information from each microphone signal through the entropy-based probability estimation.
Solution Approach 2:
The patent performs preliminary optimization of the speech presence probability estimation before applying the MVDR algorithm. By pre-optimizing the probability models using entropy calculation and iterative refinement, the system prepares high-quality input parameters that enable the MVDR algorithm to achieve good enhancement results even with limited microphone resources.
3Measurement precision
If traditional VAD-based methods are used, then the system is simple to implement, but the noise covariance matrix estimation accuracy is poor
Solution Approach 1:
The patent replaces the simple but inaccurate traditional VAD algorithm with an entropy-based probability optimization system. This substitution introduces information-theoretic principles to transform the estimation mechanism, achieving high-precision noise covariance matrix estimation by maximizing the entropy of the probability distributions through iterative optimization.
Solution Approach 2:
The patent implements a feedback mechanism where the entropy of the speech presence and absence probabilities is continuously calculated and used to guide the optimization process. The iterative algorithm adjusts the probability estimates based on entropy feedback, refining the noise covariance matrix estimation until convergence is achieved, thereby ensuring high estimation accuracy.
Data Source
AI summary
The present disclosure provides a method and system for determining a speech presence probability, a speech enhancement method and system, and a headphone. The speech presence probability and a speech absence probability in an iteration operation may be corrected by comparing an entropy of the speech presence probability and an entropy of a speech absence probability, such that a faster convergence speed and better convergence results may be obtained, thereby improving accuracy of an estimation of the speech presence probability and an accuracy of an estimation of a noise spatial covariance matrix, and then improving a speech enhancement effect of a minimum variance distortionless response (MVDR).