Multi-Mic Call Noise Reduction Using Echo-Aware Feature Fusion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional dual-mic noise reduction solutions struggle to effectively distinguish human voices from environmental noise and other voices, resulting in limited noise reduction capabilities.
Innovation Solution
A call noise reduction method utilizing complex convolutional neural networks to extract fusion features from multiple microphones' signals, followed by noise reduction processing using a convolutional neural network and prediction network to generate a noise-reduced call signal.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional dual-microphone beamforming and single-channel AI noise reduction are used, then the device complexity is low, but the noise reduction capability is limited and cannot effectively distinguish human voices from environmental noise
Solution Approach 1:
The patent segments the noise reduction task into multiple stages: first using traditional dual-mic beamforming to separate spatial signals, then applying independent single-channel AI noise reduction models to each microphone channel. This segmentation allows the system to achieve advanced noise reduction capability while maintaining manageable computational complexity by dividing the problem into smaller, independently processable parts.
2Reliability
If multiple microphones are used to improve noise reduction, then the noise reduction capability is enhanced, but the device complexity and computational load increase
Solution Approach 1:
The patent extracts and processes features from each microphone channel independently through separate AI models, then combines the results. This extraction approach allows the system to leverage multiple microphones for improved noise reduction while optimizing computational efficiency by processing each channel separately rather than handling all microphones simultaneously in a single complex model.
3Measurement precision
If complex convolutional neural networks are used to extract fusion features, then the ability to distinguish human voices from noise is improved, but the memory usage and computational load increase
Solution Approach 1:
The patent applies different processing approaches to different parts of the signal processing pipeline: traditional beamforming for spatial separation, independent single-channel AI models for each microphone, and feature fusion only where necessary. This localized application of complex algorithms only where needed maintains high voice distinction accuracy while reducing overall memory usage compared to applying complex CNNs to all signals uniformly.
Data Source
AI summary
This application discloses a call noise reduction method and devices, including earphones. The method comprises: acquiring an echo cancellation reference signal, a first noise reduction reference signal received by a first microphone, and a call signal received by a second microphone; extracting a first fusion feature of the first noise reduction reference signal and the call signal, and extracting an echo signal feature of the echo cancellation reference signal; fusing the first fusion feature and the echo signal feature to generate a combined feature; and using the combined feature to perform noise reduction processing on the call signal to generate a noise-reduced call signal.