Task for Cuetessa, Inc. – Predicting Valence of Pop Songs

Objective

Valence describes the musical positiveness of a track, e.g. ranging from sad/depressed to happy/cheerful. An automatic method of classifying the valence of pop songs is useful for playlist curation and other applications. The aim of this task is to develop a Python-based module to predict the valence of newly released pop songs. Publicly available datasets can be used for training and testing

Two approaches are to use as input:

• the audio data (e.g., .wav files) of songs
  • Regression model trained using audio features extracted from audio files
• The lyrics of songs.
  • Bidirectional Encoder Representations from Transformers (BERT) model to predict valence from lyrics

Due to lack of sufficient resources and the scale of this project:

• Sample of 500 songs taken from original dataset
• Audio previews for respective songs extracted for feature extraction

Data Description

Lyrics dataset on Kaggle that contains:

• full lyrics and labels of more than 150,000 songs
• The label is the Spotify valence attribute, ranging from 0 to 1.
  • It describes the musical positiveness conveyed by a track.
  • Tracks with high valence sound more positive (happy, cheerful), while tracks with low valence sound more negative (sad, depressed).
• Music attributes such as (tempo, danceability,loudness,etc) extracted and added to original dataset using Spotify’s Web API for Developers.
  • Spotify's Web API can extract relevant song data from the Spotify music catalog

Libraries used

- pandas
- numPy
- matplotlib
- seaborn
- spotipy
- transformers
- sklearn
- Librosa

Models Evaluated

- Support Vector Regressor
- RandomForestRegressor
- KNeighborsRegressor
- XGBRegressor
- CatBoostRegressor
- LGBMRegressor

Insights

• Further training the model based on such dataset and assess if val_MAE is improving.
• Using the valence predicted by the CNN and the valence probability predicted for the songs’ lyrics by an NLP model as inputs to train a linear regression mixed model.
• Genre classification can also be completed using CNN
• Building a larger dataset by hiring a third-party provider (such as Appen) to rate Spotify songs preview’s valence.