Speech Denoising Neural Network Using Deep Feature Losses
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing speech processing systems face challenges in effectively denoising audio signals degraded by background noise, requiring complex statistical signal processing and expert knowledge for manual tuning of loss functions.
Innovation Solution
A denoising neural network trained with deep feature losses extracted from a pre-trained audio classifier neural network, which compares internal activation patterns between noise-free and noisy signals, allowing for efficient noise reduction without manual tuning, and can be implemented in various devices and platforms.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If complex statistical signal processing is used for speech denoising, then noise reduction effectiveness is improved, but system complexity and need for expert knowledge increase
Solution Approach 1:
The patent replaces complex statistical signal processing mechanisms with a neural network-based deep learning system. The neural network automatically learns denoising patterns from training data, substituting manual statistical methods with an automated learning-based approach that reduces the need for expert knowledge while maintaining or improving denoising effectiveness.
Solution Approach 2:
The patent changes the fundamental parameters of the denoising approach by using learned features from neural network layers instead of fixed statistical parameters. The system transforms the input speech signal through multiple neural network layers, automatically adapting the processing parameters based on the learned representations rather than relying on predetermined statistical models.
2Reliability
If manual tuning of loss functions is performed, then denoising performance is improved, but time and expertise requirements increase
Solution Approach 1:
The patent implements self-service by enabling the neural network to automatically learn optimal denoising parameters and loss function weights during the training process. The system performs self-adjustment through backpropagation and gradient descent, eliminating the need for manual tuning by experts. The network autonomously optimizes its own performance metrics during training.
Solution Approach 2:
The patent applies preliminary action by pre-training the neural network on large datasets of noisy and clean speech pairs before deployment. This preliminary training phase automatically establishes the optimal loss function parameters and feature weights, so that when the system is deployed, no additional manual tuning is required - the tuning work has already been performed in advance during training.
3Manufacturing precision
If deep feature losses from multiple time scales are used, then denoising quality is improved, but computational complexity increases
Solution Approach 1:
The patent applies segmentation by dividing the speech signal processing into multiple neural network layers, each operating at different time scales and extracting different feature representations. This hierarchical segmentation allows the system to process information at multiple resolutions simultaneously, improving denoising quality while distributing computational load across specialized layers rather than requiring all computations at once.
Data Source
AI summary
Techniques are provided for speech denoising using a denoising neural network (NN) trained with deep feature losses obtained from an audio classifier NN. A methodology implementing the techniques according to an embodiment includes applying the speech denoising NN, to be trained, to a noisy sample of a training speech signal to generate a processed training speech signal. The method further includes applying a trained audio classifier NN to the processed training speech signal to generate a first set of activation features, and applying the trained audio classifier NN to a clean sample of the training speech signal to generate a second set of activation features. The method further includes calculating a loss value based on the first and second sets of activation features, and performing a back-propagation training update of the denoising NN, based on the loss value. The method includes iterating this process to further train the denoising NN.