Neural Network-Based Low-Frequency Perception Enhancement Used in Biomedical Hearing Applications

The reduction in hearing sensitivity at low frequencies is a common form of hearing loss which can significantly affect the perception. In this paper, we propose a neural network-based virtual bass system with stationary/transient source separation to address this problem. The system enhances the harmonics of the low-frequency component of the audio signal to increase the sound of the simulated low-frequency signal. We use two main approaches to virtual bass enhancement: non-linear devices (NLDs) and phase vocoders (PVs), meanwhile NLDs are more suitable for transient signals such as drums and percussion and PVs are more suitable for stationary signals such as vocals. We used a neural network to separate the signals into transient and stationary components and applied separate virtual bass enhancement to achieve a complete virtual bass system. In listening tests, our hybrid system achieved improved bass perception and lower distortion compared to previous algorithms. Our system has the potential to improve the quality of life for individuals with low-frequency hearing loss by enhancing their ability to perceive low-frequency sounds in sound and speech.