Machine Learning Noise Suppression with Uncertainty Gain Control
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing noise suppression technologies in microphone output signals face challenges in balancing noise reduction and speech distortion, often prioritizing one over the other, leading to suboptimal performance in audio communications.
Innovation Solution
A machine learning program is used to obtain multiple outputs for different frequency bands, determining an uncertainty value and gain coefficient that adjusts noise suppression based on tuning parameters to optimize speech audibility, allowing for trade-offs between noise reduction and speech distortion.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Object-affected harmful factors
If high levels of noise suppression are applied, then noise reduction is improved, but speech distortion increases
Solution Approach 1:
The patent implements dynamic adjustment of gain coefficients through a machine learning program that processes multiple frequency bands independently. The system adapts noise suppression levels in real-time based on uncertainty values and tuning parameters, allowing optimal balance between noise reduction and speech preservation that changes with input conditions
Solution Approach 2:
The patent applies different gain coefficients to different frequency bands, treating each band with locally optimized noise suppression. This allows aggressive noise reduction in frequency regions where speech is absent while maintaining gentle processing in speech-containing bands, resolving the contradiction between overall noise reduction and speech preservation
2Manufacturing precision
If noise suppression is minimized to retain desired sound, then speech distortion is reduced, but noise suppression performance decreases
Solution Approach 1:
The system dynamically switches between conservative and aggressive noise suppression strategies based on uncertainty values. When uncertainty is low (high confidence in speech presence), the system minimizes suppression to preserve speech quality. When uncertainty is high (low confidence), the system applies stronger suppression to reduce noise, optimizing the trade-off adaptively
Solution Approach 2:
The patent changes the gain coefficient parameter based on uncertainty values and tuning parameters. By adjusting this key parameter dynamically, the system can shift between preserving speech (lower gain adjustment) and suppressing noise (higher gain adjustment), resolving the contradiction through parameter optimization
3Object-affected harmful factors
If multiple outputs are obtained for different frequency bands, then noise suppression control is improved, but device complexity increases
Solution Approach 1:
The machine learning program serves multiple functions: it processes multiple frequency bands, generates multiple outputs, calculates uncertainty values, and determines optimal gain coefficients all within a single unified model. This multi-functionality achieves sophisticated noise suppression control without proportionally increasing overall system complexity
Data Source
AI summary
Examples of the disclosure enable tuning of the noise suppression in audio signals such as microphone captured signals. In examples of the disclosure, a machine learning program can be used to obtain outputs for a plurality of different frequency bands. One or more tuning parameters can also be obtained. The outputs can be processed to determine at least one uncertainty value and a gain coefficient for the plurality of different frequency bands. The at least one uncertainty value provides a measure of uncertainty for the gain coefficient, and the gain coefficient is adjusted by the at least one uncertainty value and the one or more tuning parameters. The adjusted gain coefficient is configured to be applied to a signal associated with at least one microphone output signal within the plurality of different frequency bands to control noise suppression for speech audibility.