Hi, I urgently need help with a project assignment. I will upvote for you. I have a sound pack of 14 audio files recording car sounds. I am not sure how to divide them into classes and train classifier. Here's the instruction.

Data collection: add to the Excel spreadsheet your id, username, url and the class name. This username is selected during the registration to Freesound. It can be used by others to search for your uploaded files. If the files have been submitted as a pack, then they can be also downloaded as a pack, if not, they have to be downloaded individually. For the report, keep track of the files that you have uploaded and downloaded, indicating username(s) from which you have downloaded the samples, and the amount of files per class selected.

Feature extraction: You will have to calculate some of those features from your selected audio files. Features: energy, RMS, spectrograms, log-spectrograms, mel-spectrograms, logmel-spectrograms, MFCCs, CQT spectrograms. Once a set of features have been calculated from each audio file from your dataset, you can display few of them to try to understand what do they represent and if they show different information for each of the classes. Plotting the histograms could help to visualize which feature is able to better differentiate between two classes. Histograms show the frequency distribution of an array of values, another tool are the scatter plots which are used to display clusters of data. For example, if you calculate the energy of the audio files and plot it using histograms or scatter plots, selecting different colors for each class, if the values of one class are really similar to the other class the two histograms or the two cluster will be overlapped, while if they are different, you will be able to see them separately. Note: for features with matrix shape, if you use a SVM as a classifier, where the input has to be a vector, then you have to change from matrix to vector shape. The options are to concatenate the columns and form a long vector, or do the average along the frame axis.

Define model: Train simple binary classifier. Train the model with the training data, the model learns how the data is structured and for the case of knn learns to cluster, the validation data is used to tell the model if it is leaning to separate the two classes correctly, once we fit the model then we can use the test data to finally assess how well our model performs with unseen data.