🎵 ChromaLite: Neural Network-based Musical Scale Classifier

ChromaLite is a machine learning pipeline for predicting Western musical scales from audio using chromagrams. It includes a custom dataset generator, chroma feature extraction, and a simple neural network built with PyTorch.

Think of it as the underlying system for a “Shazam for scale detection” — trained on music theory data.

---

🛠️ Tech Stack

Music and Audio: Librosa, music21, FluidSynth

---

📁 Project Highlights

• 🔧 Synthetic Dataset Generator
  Generates labeled chroma tensors for 24 Western scales using music21, librosa, and MIDI-to-audio synthesis.

• 🎼 1,900+ Samples Across 24 Scales
  Each sample contains randomized notes in a specific scale, saved as chroma tensors and scale indices.

• 🧠 Custom Neural Network Model (ChromaLite)
  A simple PyTorch-based neural network that learns to classify musical scales from [1, 12, T] chroma inputs.

---

🔬 Dataset Overview

• Input: Chroma tensor [1, 12, T] (12 pitch classes, variable time)
• Output: Integer scale index (0–23, covering major and minor scales)
• Format: .pt and .csv versions provided
• 👉 View Dataset on Kaggle

---

📈 Results

Train Accuracy	Test Accuracy
79.1%	78.28%

---

🔭 Future Work

• Add rhythmic diversity, chords, and velocity variation
• Generate more unique samples
• Build more models