Hello and welcome!
The main goal of this project was to use raw audio data to predict valence in order to build build audio recommendation software. Valence is an index of musical energy and emotion evoked by the song. We used Librosa, a Python module for audio analysis to build spectrograms. Several features were extracted and fed into machine learning algorithms such as including KNN, XGBoost and Neural Networks. Our findings revealed that annotations for streaming platforms are not good enough for a reliable model, even when using several sets of features extracted from the audio signal. The audio recommendation model will require high quality annotated data for a reliable performance, although Root Mean Square Error of 8% was achieved
Please open the PDF file in the main branch to view the report. The python code is available as Jupyter Notebook.
Thank you for checking it out!