ML Feedback Cancellation in Audio Systems
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Audio amplification systems face challenges in reducing howling sounds caused by feedback between speakers and microphones, leading to artifacts that degrade the signal-to-noise ratio and make speech unintelligible.
Innovation Solution
The implementation of machine learning-based adaptive feedback cancellation, which trains models to distinguish and remove feedback components from audio signals while preserving desirable sound components, using independent models for different frequency subbands to enhance processing efficiency and accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Object-affected harmful factors
If conventional adaptive filters are used for feedback cancellation, then the system can remove feedback components, but the desirable sound components are also removed along with the feedback, degrading audio quality
Solution Approach 1:
The patent divides the audio signal processing into multiple frequency subbands (e.g., low-frequency subband, mid-frequency subband, high-frequency subband). Each subband is processed independently by dedicated machine learning models, allowing selective feedback cancellation in each frequency range while preserving desirable sound components in other subbands.
Solution Approach 2:
Different machine learning models are trained and applied to different frequency subbands, with each model optimized for the specific characteristics of its assigned subband. This local specialization enables more effective feedback cancellation in each frequency range while preserving the unique qualities of desirable sounds in that subband.
2Object-affected harmful factors
If machine learning models process the entire frequency band, then comprehensive feedback cancellation is achieved, but computational complexity and processing time increase significantly
Solution Approach 1:
The patent segments the frequency spectrum into multiple subbands and assigns separate machine learning models to each subband. This segmentation reduces the computational burden on each individual model compared to processing the entire frequency band with a single model, while still achieving comprehensive feedback cancellation across all frequencies.
Solution Approach 2:
Instead of using one large model to process all frequencies, the patent employs multiple smaller models that each process a portion of the frequency spectrum. This partial action approach reduces the computational complexity of each model while collectively achieving complete frequency coverage through the combination of multiple models.
3Object-generated harmful factors
If aggressive noise suppression is applied to remove feedback, then feedback artifacts are reduced, but the signal-to-noise ratio of desirable speech deteriorates
Solution Approach 1:
The patent applies different processing characteristics to different frequency subbands, with each machine learning model optimized to preserve desirable sound components specific to its subband while removing feedback. This local quality approach prevents the aggressive suppression of desirable speech signals that would occur with uniform noise suppression across all frequencies.
Solution Approach 2:
The patent converts the potentially harmful effect of feedback into a detectable pattern that machine learning models can identify and remove. By training models on training data that includes feedback characteristics, the system learns to distinguish feedback from desirable sounds and removes only the harmful feedback components while preserving the beneficial speech and environmental sounds.
Data Source
AI summary
This disclosure provides systems, methods, and devices for audio signal processing that support feedback cancellation in a personal audio amplification system. In a first aspect, a method of signal processing includes receiving an input audio signal, wherein the input audio signal includes a desired audio component and a feedback component; and reducing the feedback component by applying a machine learning model to the input audio signal to determine an output audio signal. Other aspects and features are also claimed and described.