A novel noise-reduction algorithm for real-time speech processing

Abstract

We developed a new noise-reduction algorithm based on a joint spectro-temporal representation of signals. The algorithm was inspired by the discovery in our laboratory of higher-level avian auditory cortical neurons that showed invariant responses to communication signals. The algorithm consists of an analysis step and a synthesis step. In the analysis step, the sound is first decomposed into narrow band signals by a frequency filter bank. These time-frequency waveforms are then further analyzed using a spectro-temporal modulation filter bank to obtain a representation that is akin to the one generated by cortical neurons. In our algorithm, this modulation filter bank was obtained from the principal component analysis of the speech signal in the time-frequency representation. We then learned which subset of the modulation filters provided the best information to extract the signal from the noise. In the synthesis step, we then used this subset of spectral-temporal modulation feature detectors to generate a set of time-varying frequency gains that could be applied directly to the original time frequency decomposition. In this manner, we were able to perform noise reduction in real time and with minimal delay. Our algorithm yielded similar noise reduction but better quality speech quality than current state-of-the-art algorithms.