Neural Acoustic Howling Suppression With Recursive RNN Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing acoustic howling suppression methods, such as gain control, notch filters, and adaptive feedback cancellation, are inadequate in scenarios requiring high acoustic amplification, often distort the target sound, and introduce unexpected howling frequencies, necessitating a more effective solution.
Innovation Solution
A neural-network based AHS model is trained using a teacher-forcing learning approach, incorporating a Kalman filter, to suppress acoustic howling by transforming the recurrent suppression process into an instantaneous speech separation process, utilizing attention-based recurrent neural networks to extract the target signal and attenuate the playback signal.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Object-affected harmful factors
If gain control or notch filter methods are used to suppress acoustic howling, then howling is attenuated, but the target sound is distorted and speech quality deteriorates
Solution Approach 1:
The patent replaces traditional mechanical signal processing methods (gain control, notch filters, adaptive feedback cancellation) with a deep learning-based neural network system. The neural network is trained to distinguish between howling components and target speech, enabling suppression of howling while preserving speech quality through learned patterns rather than fixed filtering mechanisms.
Solution Approach 2:
The patent transforms the howling suppression problem from a fixed-parameter filtering task into a dynamic parameter estimation problem. The neural network learns to estimate time-varying parameters of howling signals (frequency, amplitude, temporal characteristics) and adapts its suppression strategy accordingly, allowing preservation of speech components that would otherwise be attenuated by fixed filters.
2Object-affected harmful factors
If adaptive feedback cancellation is used to attenuate howling, then howling is suppressed, but de-correlation techniques distort speech quality
Solution Approach 1:
The patent replaces the mechanical de-correlation technique with an intelligent neural network-based approach. Instead of forcing statistical independence between signals through mathematical transformations, the neural network learns the correlation structure between howling and speech from training data and makes intelligent decisions about which components to suppress, preserving speech information that would be distorted by de-correlation.
3Object-affected harmful factors
If traditional AHS methods are used, then howling is suppressed, but the system is inadequate for high acoustic amplification scenarios
Solution Approach 1:
The patent introduces dynamic adaptability through a neural network that can adjust its suppression strategy in real-time based on the acoustic conditions and amplification level. The system is trained on diverse scenarios including high amplification conditions, enabling it to adapt its parameters and behavior dynamically rather than relying on fixed thresholds or gain control mechanisms that fail in extreme conditions.
Solution Approach 2:
The patent enables the system to handle high amplification scenarios by training the neural network to recognize and process signals with varying amplitude ranges and signal-to-noise ratios. The network learns parameter transformations that allow it to effectively suppress howling across a wide dynamic range, making the system versatile for different amplification levels including high-gain scenarios where traditional methods fail.
Data Source
AI summary
A method and apparatus comprising computer code configured to cause a processor or processors to receive an audio signal obtained from a microphone, input the audio signal into a neural-network based AHS model, train the neural-network based AHS model based on input signals which are recursively generated from the audio signal during training of the AHS model, and output an AHS signal from the neural-network based AHS model in which AHS is applied to the audio signal, wherein the AHS signal is a version of the audio signal in which acoustic howling noise of the audio signal is suppressed and target audio of the audio signal is sustained.