Separation of vibration-derived sound signals based on fusion processing of vibration sensors and microphones

Abstract

This paper proposes a sound source separation method for vibration-derived sound signals such as sounds derived from mechanical vibrations by using vibration sensors. The proposed method is based on two assumptions. First, a vibration signal and the sound derived from the vibration are assumed to have a linear correlation. This assumption enables us to model the vibration-derived sound as a linear convolution of a transfer function and a vibration signal recorded by a vibration sensor. Second, un-vibration-derived sound signals such that the sound source is not connected to vibration sensors via a solid medium are barely recorded by vibration sensors. This assumption leads to a constraint of the transfer function from the un-vibration-derived sound sources to the vibration sensors. The proposed framework is the same as a microphone-array-based blind source separation framework, except that the proposed method constructs arrays with microphones and vibration sensors, and the separation parameters are constrained by the prior knowledge gained from the above second assumption. Experimental results indicate that the separation performance of the proposed method is superior to that of a conventional microphone-array-based source separation method.