Spectral Similarity Estimation and ML-Based Classification of Speech Using PDMS-Embedded FBG Sensor-Based Contact Microphone
Harshini Surapaneni, Archana Thrikkaikuth Chalackal, Srijith Kanakambaran · IEEE Sensors Journal · 2024
A fiber Bragg grating contact microphone (FBGCM) system for speech classification has been demonstrated in this work. Ambient noise insensitivity feature of these microphones gives them an edge over the conventional ones. When attached to the surface of an audio device or the human throat, they pick up vibrations generated by the sound produced. The captured vibrations are converted into shifts in the Bragg wavelength, enabling the extraction of spectral information. Time domain, frequency domain, volume-amplitude, noise, and similarity analyses were conducted to characterize the FBGCM, and machine learning (ML) models were used for speech classification. A nonlinear spectral response has been observed by the analysis of magnitude spectrum in the range of 100–2000 Hz. An amplitude of the recorded data increased linearly up to 60% volume level and saturated thereafter. The FBGCM also demonstrated superior noise immunity compared with a standard voice recorder (VR), with its average signal-to-noise ratio (SNR) values being 5.4 dB higher. A spectral similarity analysis indicated that the horizontal placement of FBGCM resulted in a better response compared with the vertical orientation. Furthermore, 70% of the target words showed an improvement in the spectral response with a PDMS-embedded FBG sensor. An accuracy of 90% and a precision of 91% have been achieved for speech classification using ML models.