Music_Recommendation

🎶 Music Recommendation System using LanceDB

This project implements a music recommendation system using audio feature extraction and vector similarity search. By utilizing LanceDB, PANNs for audio tagging, and Librosa for audio feature extraction, the system finds and recommends tracks with similar audio characteristics based on a query song.

📝 Project Overview

The Music Recommendation System allows users to:

• Upload their own audio files (in MP3 format) or choose from pre-defined sample tracks.
• Extract audio embeddings using a pretrained audio tagging model (PANNs).
• Perform a vector similarity search to find similar tracks based on extracted audio features.
• Play back both the query track and the recommended tracks.

🚀 Key Features

• Audio Embedding Extraction: Extracts audio embeddings using PANNs (Pretrained Audio Neural Networks).
• Vector Similarity Search: Utilizes LanceDB for fast vector similarity searches to recommend songs based on their audio embeddings.
• Interactive UI: Built using Streamlit, offering an intuitive interface to upload or select audio files and view recommendations.
• Supports Various Genres: Works with a dataset containing songs from multiple genres.

📦 Installation

To set up the project, follow these steps:

1. Clone the Repository:

   git clone https://github.com/lancedb/vectordb-recipes
   cd applications/Music_Recommandation

2. Install Required Dependencies: Install all required libraries using the following command:

   pip install -r requirements.txt

3. Set up the Pretrained Model: The pretrained PANNs model is used to extract audio embeddings. You can either:

   • Download the checkpoint automatically (handled by the app).
   • Or manually place the Cnn14_mAP=0.431.pth checkpoint in the panns_data folder.

   You can also download the model checkpoint manually from this link.

🎵 Dataset

The Ludwig Music Dataset is a compact Music Information Retrieval (MIR) dataset, sourced from Discogs and AcousticBrainZ, specifically created to support the Ludwig Backend of the SpotMyFM music recommendation engine.

Dataset Overview

• MP3 Files: Each track is represented by a 30-second fragment, with filenames corresponding to their respective Spotify Track ID. These fragments have been extracted from Spotify and organized by genre.
• MFCCs: Contains .npy files for each track, with Mel-frequency Cepstral Coefficients (MFCCs). Each .npy file holds approximately 10 MFCCs, with each segment lasting 3 seconds.
• Spectrograms: .npy files contain Mel-frequency Spectrograms for each track. Each .npy file includes about 10 spectrogram segments, each 3 seconds in duration.
• Labels Metadata: The labels.json file includes rich metadata for each track, such as genre, subgenre, and mood.
• Subgenres Metadata: The subgenres.json file contains a balanced subset of track IDs associated with each subgenre.

Folder Structure

The dataset is organized as follows:

../data/ludwig_music_data
├── mfccs
│   ├── reggae/xzxx.npy
│   └── ... (additional tracks by genre)
├── mp3
│   ├── reggae/xzxx.mp3
│   └── ... (additional tracks by genre)
├── spectrogram
│   ├── reggae/xzxx.npy
│   └── ... (additional tracks by genre)
├── subgenres.json
└── labels.json

You can download a smaller sample dataset used for this project here.

🔧 Usage

1. Run the Streamlit App: Start the Streamlit app with the following command:

   streamlit run app_music.py

2. Using the App:

   • Upload a music file (MP3 format only) or choose one of the sample tracks.
   • The app will load the selected audio, extract features, and find similar tracks from the database.
   • You can listen to both the query track and recommended tracks directly in the app.

🧠 How It Works

1. Audio Embedding Extraction:

   • The uploaded or selected song is loaded using Librosa, and features are extracted using PANNs (Pretrained Audio Neural Networks).

2. Vector Search with LanceDB:

   • Once the audio embedding is generated, it is passed to LanceDB to perform a vector similarity search. This returns the top tracks that have similar audio features.

3. Recommendations:

   • The results of the similarity search are displayed in the app, and the user can play the recommended songs.

🤝 Contributing

Contributions are welcome! Please fork this repository and submit a pull request for any features, bug fixes, or improvements.