Multistage Deep Learning Noise Suppression for Single Microphone
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing noise suppression techniques for single microphone systems are inadequate in removing noise without distorting the underlying speech signal, especially in harsh environments with non-stationary noises or low signal-to-noise ratios.
Innovation Solution
A multi-stage deep learning noise suppression system that uses AI/deep learning neural networks to estimate noise power spectra and apply gain values to reduce noise in a noisy input signal without distorting the speech signal.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Object-affected harmful factors
If traditional DSP-based noise suppression techniques are used, then noise reduction is achieved, but the speech signal is degraded and distorted
Solution Approach 1:
The patent replaces traditional DSP-based noise suppression mechanisms with deep learning neural networks. The system uses a first stage neural network to estimate noise power spectrum and generate noise gain values, then applies these gains to the audio signal. A second stage neural network estimates clean signal power spectrum and generates corresponding gains. This substitution of mechanical/DSP-based processing with AI-based processing enables superior noise reduction while preserving speech signal fidelity, directly resolving the contradiction between noise reduction effectiveness and signal distortion.
2Device complexity
If single microphone systems are used, then device complexity and cost are reduced, but noise suppression performance is inadequate
Solution Approach 1:
The patent implements a self-service approach by training neural networks to automatically estimate noise power spectrum and clean signal power spectrum from the single microphone input. The system uses the input signal itself to generate the necessary gain values through the neural networks, eliminating the need for additional microphones or complex external processing systems. This self-service capability enables effective noise suppression in single microphone configurations, resolving the contradiction between device simplicity and noise suppression performance.
3Object-affected harmful factors
If deep learning neural networks are used for noise suppression, then noise removal effectiveness is improved, but computational power and processing time are increased
Solution Approach 1:
The patent segments the noise suppression process into two distinct stages handled by separate neural networks. The first stage neural network processes the input signal to estimate noise characteristics and generate noise gain values. The second stage neural network processes the result to estimate clean signal characteristics and generate corresponding gains. This segmentation allows each network to be optimized for its specific task, improving overall noise removal effectiveness while enabling more efficient computational resource utilization compared to a single monolithic approach.
4Loss of time
If real-time processing is implemented, then latency is reduced, but processing power requirements are increased
Solution Approach 1:
The patent replaces traditional computationally intensive DSP algorithms with optimized deep learning neural networks that are specifically trained for real-time inference. The networks process audio data through efficient convolutional and recurrent operations that can be executed in real-time on modern hardware platforms. This substitution enables real-time noise suppression with reduced latency while managing processing power requirements through model optimization and hardware acceleration techniques.
Data Source
AI summary
A multi-stage noise suppression system for reducing noise components in a noisy input signal has a first stage neural network that estimates a noise power spectrum for the noisy input signal. A first set of gain values corresponding to the noise power spectrum is generated by the first stage neural network. A second stage neural network estimates clean signal power spectrum values, which are derived from an application of a second set of gain values generated as a function of the clean signal power spectrum values and a first stage reduced noise signal power spectrum values.