Sound Source Distance Estimation Using Sound Frequency Attenuation

Abstract: Estimating the distance between a sound source and receiver is an important problem. This thesis explores the possibility of estimating distance from a sound source using the attenuations regarding different frequencies over a distance. This method can be useful in cases where currently existing distance approxima- tion methods are not an option. We aimed to use an image recognition convolu- tional neural network, alongside a widely used feature extraction tool in sound analysis called Mel Frequency Cepstral Coefficients to create a model that can use frequency attenuation for sound source distance estimation. Our goal is to perform a pilot study, hoping for an accuracy within a few meters. The final model in this project has an accuracy within a few meters when tested indoors with little ambient noise. The model is limited in scope and has limited data, so the results may not be reliable beyond the scope of this thesis.