Echo Cancellation via Voiced-Unvoiced Neural Network Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing echo cancellation technologies in wireless devices often use monolithic networks that inefficiently process voiced and unvoiced components, leading to compromised speech quality due to their disparate probability distributions.
Innovation Solution
A system utilizing three neural networks: a first neural network for decomposing voiced components, a second neural network for decomposing unvoiced components, and a third neural network for merging these components to generate an echo-cancelled speech signal.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If a monolithic neural network is used to process both voiced and unvoiced components, then the device complexity is reduced, but the speech quality deteriorates due to incompatible probability distributions requiring different processing weights
Solution Approach 1:
The patent divides the monolithic neural network into separate specialized networks: a voiced component processor and an unvoiced component processor. Each network is optimized for its specific component type with dedicated weights and coefficients, resolving the conflict between simplicity and speech quality by segmenting the processing task.
Solution Approach 2:
The patent applies different processing parameters and architectures to different parts of the speech signal. Voiced components receive one type of processing while unvoiced components receive another, allowing each region of the signal to be handled with locally optimized quality rather than a uniform approach.
2Manufacturing precision
If separate neural networks are used for voiced and unvoiced components, then the speech quality improves through specialized processing, but the device complexity increases
Solution Approach 1:
By segmenting the processing into separate networks for voiced and unvoiced components, the patent achieves specialized processing that improves speech quality while managing complexity through clear functional separation rather than attempting to optimize a single monolithic network.
Solution Approach 2:
The overall echo cancellation system maintains universality by handling both voiced and unvoiced components through a coordinated multi-network architecture, allowing the system to process different speech types effectively without requiring completely separate systems.
3Ease of operation
If uniform weights and coefficients are applied in a monolithic network, then the processing is simplified, but the speech quality of at least one component is compromised
Solution Approach 1:
The patent implements local quality by assigning different weights and coefficients to voiced and unvoiced component processors. This allows each processor to be optimized for its specific component type, improving speech quality without requiring complex manual tuning of uniform parameters across the entire system.
Solution Approach 2:
The patent changes the processing parameters differently for voiced versus unvoiced components. By allowing parameter variation between component types rather than enforcing uniform parameters, the system achieves better speech quality while maintaining manageable complexity through specialized parameter sets.
Data Source
AI summary
A method includes performing, at a first neural network, a first decomposition operation on a transformed input speech signal to generate a voiced component of the transformed input speech signal. The transformed input speech signal includes frequency-domain transformed near-end speech components stacked with frequency-domain transformed far-end speech components. The method also includes performing, at a second neural network, a second decomposition operation on the transformed input speech signal to generate an unvoiced component of the transformed input speech signal. The first neural network and the second neural network perform echo cancellation on the transformed input speech signal. The method further includes merging, at a third neural network, the voiced component and the unvoiced component to generate a transformed output speech signal.