##Synopsis
This project describes an approach on Soundscape Quality Estimation. To the best of our knowledge, the proposed method provides a novel approach to this problem by introducing multi-label classification, in order to assess the quality of a soundscape (i.e. an audio landscape) based on the qualitative evaluation of its individual sound elements. To achieve this task we employ a Deep Convolutional Neural Network (CNN) which operates on pseudocolored RGB frequency-images, which represent audio segments.
The repository consists of the following modules:
- Audio segmentation using the PyAudio analysis library
- CNN training using the Lasagne Deep-Learning Framework.
- Audio classification using:
- CNNs
- CNNs using an ImageNet pre-trained model to initialize the neuron values
- CNNs using data augmentation
- An audio dataset consisting of 30 second multi-label annotated instances of soundscape auditory data. At this point the data are available in the form of spectrograms. (to be added) The instances are annotated as e.g. {vehicles, voice_(children), rain} or {sirens, shouting}.
##Installation
- Dependenices
* Installation instructions offered in detail on the above links
- Change the frequency of the audio files into 16000 Hz using changeFreq.py
- Convert your audio files into pseudocolored RGB or grayscale spectrogram images using generateSpectrograms.py
Data should be pseudo-colored RGB spectrogram images of size 227x227 as shown in Fig1 :
Fig1. - Sample RGB Spectrogram - Distribute the generated spectrograms to their respective classes using fixDataset.sh
