spack: A Simple Pipeline for Audio Classification using KAPRE

Term Project, CSC602: Machine Learning, Spring 2024, IIIT Kalyani
Name: Nikhil Laxminarayana, Harsh Singh Rawat
Reg: ECE/21136/796, ECE/21127/787

We present a pipeline for training classifiers over a diverse dataset to classify instrument samples using frequency domain feature extraction with the help of KAPRE: Keras Audio PREprocessors library. We choose to name this tool SPACK. This is also the term project for the course CSC602: Machine Learning, offered in Spring 2024 at IIIT Kalyani.

Usage and Installation

• Clone this repository, then create a virtual environment using venv or conda and then to install the dependencies type pip install -r requirements.txt in the terminal where the cloned repository resides after activating the virtual environment.
• Then run clean.py to generate a clean directory. These are the processed wavfiles used for classification. Then run train.py --model-type conv1d to train a Conv1D net over the clean dataset.
• Change the value of the model type to conv2d and then to lstm to store the model parameters into models directory.
• The models are now ready to be used to predict values. To predict values over the trainig data, run predict.py with the output logs in y_pred.npy file.
• Some notebooks in notebooks/ directory contain code snippets used to generate the confusion matrix and ROC characteristics for the models.

Plots and Metrics

We are able to achieve good accuracies over the testing data, containing diverse audio sample .wav files.

Confusion Matrix for LSTM	ROC for LSTM

Train accuracy vs epoch for the LSTM model

References

1. Choi, Keunwoo & Joo, Deokjin & Kim, Juho. (2017). Kapre: On-GPU Audio Preprocessing Layers for a Quick Implementation of Deep Neural Network Models with Keras. link