PHASE-SHIFTED VIBRATION RESPONSES OF VISCOELAS-TIC DIAPHRAGMS FOR SOUND VISUALIZATION

초록

In this study, we present initial efforts for a new sound visualization approach aimed at producing different spec-trogram-like training images for convolutional neural network (CNN)-based speech recognition. We explored the potential of the viscoelastic membrane (diaphragms) to deliver audio visualization images from acoustic sound to binarized time series and it can be visualized. The responses of the viscoelastic diaphragm are measured in two different types of data. The Lissajous curve is the first and it is physically obvious that sound input changed into phase-shifted two waves by its pattern. Although the Lissajous curve is obvious, it is improper to visualize sound as an image for future application to the deep learning algorithm, so the phase shift between two binarized vibra-tions on the viscoelastic diaphragm is measured by two laser Doppler vibrometers. Displacement of the transverse and longitudinal direction of diaphragm excitation is measured by laser reflection on a T-shaped plate with negli-gible mass. However, a bulky and expensive laser vibrometer is not suitable for future speech recognition appli-cations. Thus, responses to binarized vibration on the viscoelastic diaphragm are measured again with a triaxial accelerometer. Binarization of input sound pressure on the viscoelastic diaphragm is confirmed by comparing measured displacements and the similarity of their Lissajous curves with a laser vibrometer and triaxial accelerom-eter. Herein, we report that the proposed sound visualization method will be easily applicated to future speech recognition by combining two phase-shifted vibration responses of the viscoelastic diaphragm because the re-sponses are time variant which has high versatility on sound visualization.