Classification of Room Characteristics Using Convolutional Neural Networks: Machine Learning and Room Acoustics
Translated title
Convolutional Neural Networks Anvendt Til Klassificering af Rumkarakteristikker
Authors
Nielsen, Sanne Damhus ; Nielsen, Morten Østergaard
Term
4. semester
Education
Publication year
2019
Submitted on
2019-06-07
Pages
101
Abstract
At forstå, hvordan rum påvirker lyd, er vigtigt i blandt andet udligning af rumrespons og retsakustik. I denne kandidatafhandling undersøges, om konvolutionelle neurale netværk kan klassificere to rumegenskaber ud fra rumoverføringsfunktioner (RTF’er) – målinger, der beskriver, hvordan lyd ændrer sig fra en kilde til en mikrofon i et rum. De to egenskaber er rumvolumen og vægadmittans (et mål for hvor let vægge reagerer på og absorberer lyd). Modellerne trænes med superviseret læring på simulerede RTF’er fra en fysikbaseret lydfeltmodel afledt af Helmholtz-ligningen for bølgeudbredelse i små rum. Resultaterne viser, at modellerne klarer sig bedre, når mikrofoner placeres i regelmæssige gitter i stedet for tilfældigt. Ved klassifikation af rumvolumen opnås over ca. 90% nøjagtighed, når mere end én mikrofon bruges. For vægadmittans ligger nøjagtigheden over 80%. Modellerne til rumvolumen er mere følsomme over for additiv støj end vægadmittans-modellerne, som stadig leverer bedre end initial nøjagtighed ved 0 dB SNR.
Understanding how rooms affect sound matters for tasks like room response equalization and forensic audio. This thesis investigates whether convolutional neural networks can classify two room properties from room transfer functions (RTFs)—measurements that describe how sound changes from a source to a microphone inside a room. The two targets are room volume and wall admittance (a measure of how readily walls respond to and absorb sound). The models are trained with supervised learning on simulated RTFs generated by a physics-based sound field model derived from the Helmholtz equation for wave propagation in small rooms. Results show that performance improves when microphones are placed on regular grids rather than randomly. For room volume classification, accuracy exceeds about 90% when more than one microphone is used. For wall admittance, accuracy is above 80%. The room volume models are more sensitive to additive noise than the wall admittance models, which still achieve better-than-initial accuracy at 0 dB SNR.
[This abstract was generated with the help of AI]
Keywords
Documents