JoaoESmoreira

The following technologies were used:

• Python 3
• Librosa — for audio processing and feature extraction.
• NumPy — for numerical computation and array manipulation.
• SciPy — for statistical analysis (skewness, kurtosis, etc.).
• FFmpeg — for reading .mp3 files.

The main goal is to implement a system capable of recommending similar songs based on their audio content. The project involves:

• Feature extraction using the Librosa framework.
• Implementation of multiple similarity metrics.
• Evaluation of recommendation quality through both objective and subjective methods.

The system uses the 4Q Audio Emotion Dataset, which contains:

• Audio excerpts in .mp3 format (including 4 query tracks).
• Metadata file panda_dataset_taffc_metadata.csv, containing columns:
  • Song, Artist, Quadrant, MoodsStrSplit, and GenresStr.
• Pre-extracted features:
  • top100_features.csv (audio feature vectors)
  • features.csv (feature descriptions)

Dataset available at: http://mir.dei.uc.pt/downloads.html

1. Preparation
   • Familiarization with real-world recommendation systems (Spotify, Jango, Last.fm, etc.).
   • Downloading and analyzing the dataset.
   • Studying the Librosa framework and the base code (mrs.py).
2. Feature Extraction
   • Processing existing features from top100_features.csv.
   • Extracting new features from the dataset using Librosa:
     • Spectral features: MFCC, spectral centroid, bandwidth, contrast, flatness, rolloff.
     • Temporal features: RMS energy, Zero Crossing Rate, F0.
     • Other features: Tempo.
   • Computing seven statistical measures for each feature:
     • Mean, standard deviation, skewness, kurtosis, median, maximum, and minimum.
   • Normalization of features to the [0, 1] range.
   • (Optional bonus) Implementing feature extraction manually using only NumPy and SciPy.
3. Similarity Metrics
   • Implementation of three distance metrics:
     • Euclidean Distance
     • Manhattan Distance
     • Cosine Similarity
   • Generating similarity matrices (900×900) and recommendation rankings for the four query tracks.
4. Evaluation
   • Objective Evaluation: comparing recommendations against metadata (artist, genre, emotion, quadrant).
   • Subjective Evaluation: human evaluation using a Likert scale (1–5) to rate recommendation quality.
   • Computing Precision scores to assess performance in both evaluation types.