An Audio Frequency Unfolding Framework for Ultra-Low Sampling Rate Sensors

Recent audio super-resolution works have achieved significant success in promoting audio quality by improving a sensor's sampling rate, e.g., from 8 kHz to 48 kHz. However, these works fail to maintain the performance when the sampling rate at the sensor is ultra-low, where the audios suffer serious frequency aliasing. In this paper, we propose an audio frequency unfolding framework that efficiently reconstructs the aliasing audios to be perceptually recognizable. The intuition is that the audios generated by humans have a regular pattern on the spectrums; by learning such a regular pattern, our framework can reconstruct audio that sounds similar to real human voices. We evaluate our framework in a perceptual way: an automatic speech recognition (ASR) system is used to judge whether the words in the reconstructed audios can be correctly recognized. In the implementation based on AudioMNIST, when reconstructing the sampling rate from 2 kHz to 16 kHz, the recognition accuracy of the reconstructed audio reaches 77.1%.