Pleasantness assessment of electric vehicle interior sounds using a long short-term memory model

Abstract

The electrification of the powertrain strongly reduces the sound pressure level in the vehicle interior. However, the absence of a broadband noise originating from the combustion engine unmasks high-frequency tonal components of the electric motor and the gearbox. As they are commonly audible in transient driving conditions, the related psychoacoustic parameters and their influence on the pleasantness have to be considered dynamically. This study presents a sequence-to-one regression model on the basis of a long short-term memory (LSTM) neural network, which models the relation between times series of psychoacoustic parameters and the overall (i.e., single-valued) pleasantness. The advantage of LSTM-based models is that they consider interdependencies between time steps. The data set to be modelled mainly consisted of pleasantness ratings of recorded sounds from the vehicle interior of pure-electric and hybrid vehicles. The data set also includes pleasantness ratings of altered sounds, where synthetic components were added, amplified or attenuated using a sound separation algorithm. The model is highly accurate in predicting the data. Thus, the LSTM model can be used as an automated pleasantness assessment for the interior sound of vehicles with electrified drives.