Blind Estimation of Room Acoustic Parameters and Speech Transmission Index using MTF-based CNNs
Suradej Duangpummet, Jessada Karnjana, Waree Kongprawechnon, Masashi Unoki · 2021 29th European Signal Processing Conference (EUSIPCO) · 2021
Room acoustic parameters, such as reverberation time$(T_{60})$and clarity$(C_{80})$, as well as the speech transmission index (STI), are essential in acoustics. However, such parameters and STI are difficult to obtain in everyday places where people exist. Blind estimation for those parameters without measuring room impulse response (RIR) is necessary and challenging. This paper proposes a method based on the modulation transfer function (MTF) and Schroeder's RIR model for estimating$T_{60}\mathrm{s}$in seven-octave bands. The estimated$T_{60}\mathrm{s}$are used to approximate the MTF and RIR. Consequently, the STI and five room-acoustic parameters, including$T_{60}$, early decay time (EDT),$C_{80}$, Deutlichkeit$(D_{50})$, and center time$(T_{s})$, can be estimated. We deploy convolutional neural networks for mapping temporal amplitude envelopes of a reverberated speech signal to$T_{60}\mathrm{s}$for the sub-bands. Simulations were carried out by estimating the five parameters and STI from unseen reverberated speech signals. The root-mean-square errors between ground-truths and estimated parameters suggest that the accuracy of the estimated$T_{60}$and STIs can be improved by about 40% and 25% compared with previous methods, respectively. The other parameters were also correctly estimated, and they are comparable with those obtained from standard measurements.