Speech Enhancement Model With Res2Net Skip Connections

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Traditional speech enhancement systems struggle with non-stationary noise and low signal-to-noise-ratio conditions, leading to less accurate noise reduction and speech quality enhancement.

Innovation Solution

A machine learning model architecture for speech enhancement that includes an encoder with residual network blocks and squeeze-excitation layers, a bottleneck with uni-directional GRU layers, and a decoder with skip connections, which processes speech waveforms to enhance speech quality.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If traditional speech enhancement systems are used, then the system structure is simple, but the performance degrades under non-stationary noise and low SNR conditions

Engineering Contradiction:

Improvespeech enhancement performanceVSAvoidsystem complexity

Core Design Contradiction:

ReliabilityVSDevice complexity

Solution Approach 1:

The patent replaces traditional mechanical signal processing methods with a machine learning-based system that uses neural networks to model and remove noise. The system trains on paired clean and noisy speech data to learn optimal noise removal patterns, achieving superior performance in non-stationary noise conditions compared to conventional approaches.

Inventive Principle:

Principle #28Mechanics substitution (Replace mechanical system)

Solution Approach 2:

The system dynamically adjusts processing parameters based on the characteristics of the input signal. By analyzing the noise profile and signal characteristics in real-time, the system adapts its enhancement parameters to optimize performance under varying noise conditions, maintaining high reliability across different scenarios.

Inventive Principle:

Principle #35Parameter changes

2Reliability

If complex machine learning models are used, then speech enhancement performance improves, but computational complexity increases

Engineering Contradiction:

Improvenoise reduction accuracyVSAvoidcomputational complexity

Core Design Contradiction:

ReliabilityVSDevice complexity

Solution Approach 1:

The patent divides the speech enhancement task into multiple processing stages, each handled by specialized neural network modules. The system segments the noise removal process into distinct operations that can be optimized independently, reducing overall computational burden while maintaining high accuracy through coordinated processing across stages.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The system applies selective processing intensity to different parts of the signal. By identifying and treating only the most problematic noise components with high computational effort, while applying lighter processing to cleaner signal regions, the system achieves high noise reduction accuracy with reduced overall computational complexity.

Inventive Principle:

Principle #16Partial or excessive action

Data Source

PatentUS20250078853A1Machine learning model architecture for speech enhancement system

Publication Date: 2025.03.06 INFINEON TECHNOLOGIES AMERICAS CORP

AI summary

Methods and systems for a machine learning model architecture for speech enhancement system. The disclosed machine learning model architecture includes, among other things, an encoder, a decoder, and a bottleneck disposed between the encoder and the decoder. The encoder includes a plurality of encoder layers, and the decoder includes a plurality of decoder layers. Each encoder layer is connected to a corresponding decoder layer via a skip connection. Each encoder layer includes a Res2Net and a squeeze-and-excitation (SE) block. The bottleneck includes a first gated recurrent unit (GRU) layers and a second GRU layer.