- Audio feature extraction and classification with the ECS-10 data set audio dataset
- ECS-10 audio data is included. It consists of 10 classes of different environmental sounds (sea waves, kids playing, etc.)
- The main goal is to compare classification accuracies for the 6 tested classifiers.
- Anaconda 2 with Python 2.7. (Python 3.6 not tested yet)
- Librosa (audio loading, audio visualization and feature extraction)
- Sci-kit learn
- Keras (Theano backend)
- Numpy, Matplotlib
- Pandas (data visualization)
A Jupyter Notebook (Python 2.7 Kernel) is added to illustrate the workflow.
The scripts for feature extraction and classification have been added as
.py files and are all loaded in the Jupyter Notebook sequentally.
Running feature_extraction.py creates a numpy array for features (feature.npy) and one for labels (label.npy).
These files will be saved in the current directory.
- MFCC
- Chroma
- Mel spectrogram
- Tonal centroid feature
- Spectral contrast
- Support Vector Machine (SVM)
- Random Forest (RF)
- Naive Bayes (NB)
- Convolutional Neural Network (CNN)
- Multilayer Perceptron (MLP)
- Recurrent Neural Network (RNN)
Note: Direct comparison between classifiers can't be donde yet since their parameters haven't been tuned to optimize accuracy yet. Out of 400 audio samples, the test set consisted on the 33% of this.
- SVM: 81.7%
- RF: 80%
- NB: 69.7%
- CNN: 71.25% (100 epochs)
- MLP: 63.125 (100 epochs)
- RNN: 66% (100 epochs)
- Compute other features: MFCC + ZCR features improve classification accuracy for speech, noise and music labels. See if it also works for the 10 classes.
- Tune optimization hyperparameters (for every classifier): Weight initialization, decaying learning rate.
- Data scaling and feature normalization (MFCC)