Research and Implementation of a Hearing Aid Based on a Mel-Phase-Spectrum-Preprocessed GAN Model

초록

This paper introduces the MPSP(Mel-Phase-Spectrum-Preprocessed) Algorithm for optimizing audio preprocessing, combined with a GAN(Generative Adversarial Network) model to enhance audio output. The MPSP replaces the phase estimation method of the Griffin-Lim algorithm with a phase pre-storing technique, thereby improving noise reduction for audio output from a low-cost, custom-built hearing aid. Experimental results demonstrate that MPSP improves SDR(source-to-distortion ratio) performance by 2.31 times compared to the Griffin-Lim algorithm. The GAN model trained on data preprocessed with the MPSP algorithm was tested in three different environments, showing superior MSE(Mean Square Error) performance over the Spectral Gating-based noise-reduce method and the Denoising Autoencoder model. In PESQ(Perceptual evaluation of speech quality) evaluations, the GAN model maintained high performance in complex environments such as classrooms and workplaces, except in extremely noisy settings like restaurants. The hearing aid employs a deep neural network model to achieve cost-effective audio noise reduction, significantly improving the quality of life for individuals with hearing impairments. By pre-training the model for deployment on embedded systems, this solution can be widely applied across various industries.

키워드