Hybrid Acoustic Howling Suppression With Kalman Filter and Deep Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing acoustic howling suppression methods face challenges in balancing suppression performance and signal distortion, particularly due to the mismatch between offline training and streaming inference in deep learning models, and the sensitivity to nonlinear distortions in audio systems.
Innovation Solution
A hybrid approach combining a frequency domain Kalman filter (FDKF) with a self-attentive recurrent neural network (SARNN) is used, where the pre-processed output from FDKF serves as an additional input for SARNN training, and the model is trained in a teacher-forced manner to convert howling suppression into a speech separation problem, with the output used to update FDKF parameters.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep learning models are trained offline and then used for streaming inference, then the model can achieve good suppression performance in training conditions, but the performance degrades in streaming inference due to mismatch between training and inference conditions
Solution Approach 1:
The patent applies preliminary action by pre-processing the input signal through a frequency domain Kalman filter before feeding it to the deep neural network. This pre-processing step prepares the signal in advance to account for potential nonlinear distortions, ensuring that the DNN receives pre-conditioned input that bridges the gap between offline training conditions and streaming inference conditions, thereby maintaining suppression performance across both scenarios
Solution Approach 2:
The patent implements feedback by using the output from the deep neural network to update the parameters of the frequency domain Kalman filter. This creates a closed-loop system where the DNN's suppression results feed back into the filter, allowing the filter to adapt its parameters based on actual streaming inference performance, thus improving robustness to nonlinear distortions while maintaining suppression effectiveness
2Ease of operation
If traditional adaptive filtering is used to dynamically adjust signals in real-time, then speech quality is improved, but the method becomes sensitive to control parameters and interferences and fails to address nonlinear distortions
Solution Approach 1:
The patent introduces an intermediary approach by combining the frequency domain Kalman filter with the deep neural network in a hybrid architecture. The Kalman filter serves as an intermediary pre-processor that handles linear components of the signal, while the DNN focuses on suppressing nonlinear distortions. This division of labor reduces the sensitivity to control parameters compared to pure adaptive filtering, as the DNN learns optimal suppression strategies without requiring manual parameter tuning
Data Source
AI summary
Method, apparatus, and non-transitory storage medium for hybrid acoustic howling suppression based on a frequency filter model and a deep neural network are provided. The method may include receiving a speech signal, the speech signal including target speech, feedback, and noise, and inputting the speech signal into a trained hybrid neural-network based howling suppression model, wherein the trained hybrid neural-network based howling suppression model is trained using training speech signal and pre-processed acoustic feedback from a first frequency filter model. The method may also include generating an enhanced speech signal with suppressed howling as an output of the trained hybrid neural-network based howling suppression model, wherein the enhanced speech signal is used to update parameters of the first frequency filter model.