spcup_2022

Matlab version used: R2021b
Python version used: 3.8.10

git clone https://github.com/Fatimanaz/spcup_2022.git 
cd spcup_2022

Now create a virtual environment (optional)

pip install -r requirements.txt

Directory Structure

1. data:
   folder consists of raw audio files and their corresponding labels.csv files. Both the training dataset as well as the evaluation dataset folders are available in the data folder.
2. MATLAB_SCRIPTS:
   folder contains the matlab scripts required to augment the original dataset. The Scripts produce three new folders in the data folder: noise_added, reverb_added, and compressed.
3. Results:
   consists of the evaluated results for eval_dataset_part1 and eval_dataset_part2 and trained models.
4. feature_extraction.py:
   python script that processes the raw audio files provided in data folder and produces handcrafted-features mfcc, and chroma_cqt features.
5. ExtractedFeatures:
   consists of mfcc_features and chroma_cqt_features extracted using feature_extraction python script
6. train.py: python script to train the model. Takes input in form of handcrafted features stored in ExtractedFeatures folder
7. eval.py: python script to produce labels on evaluation dataset

Using the trained model to evaluate results from already Extracted Features

We have already provided the extracted features (mfccs and croma cqt) csv files for all the files in the ExtractedFeatures folder.

python3 eval.py

This produces the required csv files in the following format:

	track	algorithm
0	audio_file_name	label(int)

Training the entire model from scratch

Structure of the data folder

You can look at the detailed directory structure in directory_structure.txt

• To train the entire model from scratch, raw audio files need to be augmented, processed, converted into handcrafted features and then stored in the ExtractedFeatures folder.
• Keep all the audio files in a folder. Rename the folder to spcup_2022_training_part1. Place this folder inside the data folder.
• Run adapted_script.mlx using matlab. All relative paths have already been set inside the script.
• Now just run the provided python scripts, that will generate features, train the model and store it, and evaluate the results on the evaulation dataset.

python3 feature_extraction.py
python3 train.py
python3 eval.py

---