Project Emotify

Table of Contents

• Team Members
• Development Workflow
• Setup & Running
• Data Strategy & Model Training
  • Baseline Feature Extraction (MTG-Jamendo + MERT)
  • Key Resources & Technologies
• Frontend Application

Team Members

• Myroslav Natalchenko
• Kiryl Sankouski
• Michał Zach

Development Workflow on GitHub

To ensure code stability and minimize merge conflicts, we will strictly follow a Fork & Branch workflow.

1. Each team member must fork the main Emotify repository to their personal GitHub account
2. Create a specific branch in your fork for your tasks
3. Once task is complete, open a Pull Request (PR) from your fork's branch to the upstream repository's main branch

Setup & Running

Prerequisites

• Python 3.10+
• Node.js 18+

1. Backend

# Create and activate a virtual environment
python -m venv .venv
source .venv/bin/activate       # Windows: .venv\Scripts\activate

# Install dependencies
pip install -r requirements.txt

Create a .env file in the project root with your Spotify credentials:

SPOTIFY_CLIENT_ID=your_client_id
SPOTIFY_CLIENT_SECRET=your_client_secret
SPOTIFY_REDIRECT_URI=http://127.0.0.1:5000/callback

# Start the Flask server — http://127.0.0.1:5000
python app.py

On first run the backend downloads the MERT model from HuggingFace (~375 MB).

2. Frontend

cd frontend
npm install
npm run dev   # http://localhost:3000

---

Data Strategy & Model Training

To achieve accurate and scalable emotion recognition in music, Emotify adopts a feature-based, two-step pipeline:

1. High-level audio representation extraction using a large pretrained music model
2. Supervised training of a lightweight emotion classifier on extracted embeddings

This approach allows us to decouple heavy audio processing from model training, significantly reducing training cost and improving experimentation speed.

Baseline Feature Extraction (MTG-Jamendo + MERT)

As a foundation for emotion modeling, we use the MTG-Jamendo Dataset, specifically the subset annotated with mood/theme tags.

To transform raw audio into meaningful numerical representations, we employ the pretrained MERT (Music Embedding Representation from Transformers) m-a-p/MERT-v1-95M from HuggingFace model. Each track is converted into a fixed-size embedding tensor, which is stored as a .npy file.

Model Training

Our emotion prediction model is trained directly on the extracted MERT embeddings, rather than raw audio or spectrograms.

This design provides:

• Faster training cycles
• Lower hardware requirements
• Strong generalization thanks to MERT pretraining

Key Resources & Technologies

Datasets

• MTG-Jamendo Dataset (mood/theme subset)

Pretrained Models

• MERT: https://huggingface.co/m-a-p/MERT-v1-95M

Core Stack

• Python 3.10+, PyTorch, NumPy, librosa
• Hugging Face Transformers (m-a-p/MERT-v1-95M)
• Flask + Flask-SQLAlchemy (REST API + SQLite analysis cache)
• Next.js 16 / React 19, Tailwind CSS, Recharts (frontend)

Frontend Application

The Emotify frontend is implemented as a modern web application using Next.js.