Adaptive Noise Cancellation in Smart Hearing Aids Using Reinforcement Learning

Abstract

Common noise cancellation algorithms used in current hearing aids,including spectral subtraction, Wiener filtering, and traditional adaptive filtering are prone to performance degradation in non-stationary and dynamically varying speech environments and certainly be in the real world situation which could be a noisy street, a public bus or even at a party. Such approaches are based on predetermined values of adaptation parameters or trained offline, and, therefore, are not able to react efficiently to unpredictable noise properties. In overcoming these impediments, this paper develops an adaptive noise cancellation (ANC) system based on reinforcement learning (RL) that performs continuous, context-aware, real-time noise mitigation in smart hearing aids. The proposed system that involves an RL agent interacting with acoustic environment and being told whether speech clarity and listening comfort improves should make it possible to optimize the approach to noise suppression via trial-and-error learning. A Deep Q-Network (DQN) enables the decision-making process that dynamically updates ANC filter parameters based on a concise state representation based on time frequency features via short-time Fourier transform (STFT), such as, Mel-frequency cepstral coefficients (MFCCs), instantaneous signal-to-noise ratio (SNR), and spectral flatness measures. The reward is a combination of improvements in SNR and perceptual improvements in speech quality (evaluated as perceptual evaluation of speech quality or PESQ), such that the algorithm maximises intelligibility without causing unreasonable distortion. The CHiME-4 noisy speech was used to conduct simulation experiments consisting of real background noise with the case of a street, a cafe, and a transit location. Comparison with Wiener filter, spectral subtractions and a deep speech enhancement baseline that uses a convolutional neural network also illustrates that the proposed RL-based ANC framework improves average SNR with 4.7 dB and STOI by 12.5 percent throughout various noises. These findings indicate the versatility and versatility as well as the possibility of the framework to be individualized depending on the users and hence it fits well as a candidate in future hearing aids that tend to maximize the hearing experience in the most dynamic acoustic tasks.