Final Thesis

Auditory evaluation of pinna geometry modifications

Bachelor Thesis of Schneider, Jan-Niklas

In order to provide a diagnosis about the urban sound environment, especially for traÿc, sound classification using deep learning with convolutional neural networks has received increasing attention. There are already some attempts of deep convolutional neural networks (CNN) for environmental sound classification and data augmentation. These CNNs are using spectro-temporal patterns as its input to identify or classify a given sound by recognizing features from these patterns. However, to train these networks, a lot of data is required. To solve this problem data augmentation is used, which is a method to recreate one or more existing training samples to receive additional training data. In this thesis a state-of-the-art audio data augmentation technique with help of a virtual reality environment is introduced. A synthesized sound signal is designed using an existing sound data set of vehicle pass-by sounds with combustion engines. The synthesized signal is generated by using spectral granular synthesis. This signal is implemented in a virtual reality environment and generates the desired training data set (virtual pass-by sounds) for di˙erent neural networks (NNs), like CNNs, recurrent neural networks (RNNs) and convolutional recurrent neural networks (CRNNs). Furthermore, audio descriptors like the mel frequency cepstral coeÿcients (MFCCs) and the frequency spectrum are extracted from the generated data set. This information is used as the input for the NNs. Finally these NNs are evaluated with the help of the original recordings and predictions are made with completely untouched recordings (not used for training or validation). In addition, a case study of the model’s applicability was conducted using real-world measurements. The models are examined in two di˙erent ways: first, a vehicle is inspected at di˙erent speeds and second several vehicles are examined at the same speed.