Perceptual Spectral Gain Control for Speech-in-Noise Enhancement
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing speech enhancement techniques are not robust under high noise levels and often attenuate the speech component, leading to poor performance in automatic speech recognition (ASR) systems, especially in complex noise conditions.
Innovation Solution
A signal enhancer that splits audio signals into spectral components, designates them as either wanted or unwanted based on power estimates, and applies an adaptive gain to the unwanted components to improve the relative power of the wanted component, considering perceptual thresholds to enhance intelligibility.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Object-affected harmful factors
If adaptive filtering is used to maximize signal-to-noise ratio, then noise suppression is improved, but speech component is attenuated resulting in poor ASR performance
Solution Approach 1:
The patent applies different processing strategies to different spectral components based on their characteristics. Speech components are preserved with minimal attenuation while noise components are suppressed, achieving local optimization rather than uniform processing. This resolves the contradiction by making the filtering adaptive to the local spectral properties at each frequency bin.
Solution Approach 2:
The invention dynamically adjusts filtering parameters based on the estimated speech and noise power spectra. By changing the filter characteristics adaptively according to the instantaneous signal conditions, the system can suppress noise when speech is absent while preserving speech components when present, thus improving both noise suppression and ASR performance.
2Object-affected harmful factors
If strong processing is applied to suppress noise, then noise control is improved, but speech component is attenuated leading to poor intelligibility
Solution Approach 1:
The system uses feedback from the estimated speech and noise power spectra to continuously adjust the filtering strength. By monitoring the signal characteristics and adapting the processing accordingly, the system avoids excessive attenuation of speech components while maintaining effective noise suppression, thus preserving speech intelligibility.
Solution Approach 2:
The filtering approach is made dynamic and adaptive rather than static. The filter characteristics change in real-time based on the estimated signal conditions, allowing the system to apply strong noise suppression when appropriate while automatically reducing attenuation when speech components are present, thereby maintaining intelligibility.
3Device complexity
If binary mask is used to mask noise components, then noise masking is simplified, but spectrum holes are created reducing signal quality
Solution Approach 1:
Instead of using a fixed binary mask, the invention employs adaptive gain factors that are continuously adjusted based on the estimated speech and noise power spectra. This allows for smooth transitions and avoids the creation of spectrum holes, maintaining signal quality while keeping the computational complexity manageable through efficient spectral estimation.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A signal enhancer comprises an input configured to receive an audio signal that has a wanted component and an unwanted component. It also comprises a perception analyser that is configured to split the audio signal into a plurality of spectral components. The perception analyser is also configured to, for each spectral component, designate that spectral component as belonging to the wanted component or the unwanted component in dependence on a power estimate associated with that spectral component. If a spectral component is designated as belonging to the unwanted component, the perception analyser is configured to adjust its power by applying an adaptive gain to that spectral component, wherein the adaptive gain is selected in dependence on how perceptible the spectral component is expected to be to a user. This improves the intelligibility of the wanted component.