Acoustic Scenes Classification

This project shares the python code used for the challenge DCASE 2018 task 1.

The goal of Acoustic Scene Classification is to classify a test recording into one of the provided predefined classes that characterizes the environment in which it was recorded

In this project the audio signal in input is preprocessed into a spectrogram by short-time Fourier transform SFTF. A methodology of supervised machine learning based on Convolutional Neuronal Network (CNN) classified the signal.

Struture

The program main.py launches the training.

The directory Models contains the CNN modules:

• simpleCNN: simple module to test the process
• simpleCNN_v2: same with a standard conv: conv2D + batchNorm + Relu
• CNN_Dorfer: CNN proposed by CP-JKU team to the DCASE Challenge 2018 Task 1 Subtask A (without the Gaussian Noise)
• CNN_Dorfer2: CNN proposed by CP-JKU team to the DCASE Challenge 2018 Task 1 Subtask A (with the Gaussian Noise)
• Xception: Xception modele with variable number of Middle Flow and simplified Exit Flow (one conv2d and average pooling)

The directory Data contains:

• the split of train and validation/test as in the challenge DCASE 2018 (fold2_train.txt and fold2_evaluate.txt)
• the split of initial train in train and validation (fold3_train.txt and fold3_val.txt)
• mean and standard deviation on train data (mean_train.npy and std_train.npy)

The directory Utils contains:

• audio_to_spectrogram: Convert audio signal to spectrogram with Librosa librairy
• progressbar.py: a simple progress to follow the evolution of training process
• save_model: Save a model in a file
• unzip: extract the zip file of input data