TransFiLM: An Efficient and Lightweight Audio Enhancement Network for Low-Cost Wearable Sensors
Shilin Chen, Jianhao Weng, Shicong Hong, Yanbo He, Yongpan Zou, Kaishun Wu · 2024
Wearable devices such as headphones are increasingly popular in people's lives, and there is an increasing focus on how to achieve continuous and reliable information input using these devices. However, due to constraints in computing power, low power consumption, and low operating frequencies, such devices often record and transmit signals at lower sampling rates, the resultant lower-quality signals often have catastrophic implications for system performance. Efficient real-time conversion of low-resolution speech signals to full-resolution high-quality signals using low-cost wearable sensors on edge devices is a challenging research endeavor. To address this, this paper designs TransFiLM, a mobile deep learning network. It allows users to obtain full-resolution high-quality audio signals using low-cost wearable sensors on edge devices. TransFiLM integrates residual learning and super-resolution networks and employs effective signal processing strategies to achieve audio upscaling and noise reduction, significantly improving audio quality. We implement a prototype on commercial devices and conduct a series of experiments to evaluate its performance. Using signal-to-noise ratio (SNR) and log-spectral distance (LSD) as evaluation metrics, TransFiLM exhibits superior performance compared to other time-domain methods in cross-user, cross-corpus, and cross-noise environment testing. Additionally, our TransFiLM network handles 8192 samples with a response time of 181 ms, which meets the requirement to run in real-time on edge devices.