Neural network-based selective noise control with sound source separation

Abstract

In noisy environments such as factories or construction sites, noise should be reduced and the speech of workers moving in the environments should be recognized. Noise control is a method for reducing existing sounds by producing a secondary sound with the same amplitude but with antiphase, and effective on low frequency noise reduction. However, conventional noise control systems reduce all existing sounds but cannot preserve a specific target sound such as speech in noise. In this study, we propose a novel system which selectively reduces noise and preserves target sound by combining a noise control system with a sound source separation system using multilayer neural networks. The sound source separation part of the system was designed to separate mixture into noise for reduction and target sound for preservation and the use as a teaching signal of supervised learning of the neural networks. The noise control part of the system was designed to reduce noise at fixed and moving evaluation points. The whole system was designed to realize high noise reduction performance preserving target sound and low computational cost for real-time control in experiments. The performance of the proposed system was experimentally verified under the conditions that the noise and target waveforms were set as both sine waves, both superposed sine waves, and band noise and a sine wave, respectively. The results indicate that the proposed system preserves target sound to some extent and reduces noise by more than 10 dB and approximately 5–6 dB at the fixed and moving evaluation points, respectively.

Keywords

• noise cotrol
• sound source separation
• blind source separation
• neural network
• network design
• experiment
• time-variant system
• acoustic environment