AI-Powered Maestro Finder

This project is a part of Neural Networks and Deep Learning (AAI-511-02) course in the Applied Artificial Intelligence Master Program at the University of San Diego (USD).

-- Project Status: Completed

Introduction

AI-Powered Maestro Finder turns symbolic music into quick, explainable predictions. It’s a complete pipeline—from raw MIDIs to trained models to an interactive Streamlit app—built to be readable, reproducible, and easy to extend.

Objective

• Make composer recognition accessible to students, musicians, and curious listeners.
• Demonstrate an end-to-end ML workflow: data wrangling → augmentation → feature extraction → modeling → friendly UI.
• Provide a clean template you can extend with new composers or tasks (style, period, instrument).

Features

• Upload MIDI or record audio (auto-transcribed to MIDI with Basic Pitch).
• Confidence bars and piano-roll visualization.
• Optional sheet-music view (MusicXML rendered via OpenSheetMusicDisplay).
• On-device inference using a compact CNN.
• Clear, modular code: utils/ for I/O, features, inference, and visualization.

Methods

Stage	Key steps
Data wrangling	De-duplicate and clean raw MIDIs; trim to piano range (A0–C8); remove empty or corrupted tracks.
Class balancing	Pitch-shift, time-stretch, and velocity-jitter minority classes to balance counts.
Feature extraction	Two parallel views: (1) Piano-roll 88×512 @ 8 fps for CNN, (2) Event tokens for LSTM (used for research and comparison).
Modeling	Train CNN on piano-rolls and LSTM on sequences; evaluate both.
Final selection	CNN chosen for the app based on accuracy, stability, and speed.
Inference	MIDI → piano-roll → CNN → probabilities. For audio, run Basic Pitch → MIDI → same pipeline.
Front end	Streamlit UI with confidence bars, piano-roll, and optional sheet music.

Results (test split)

Model	Input window	Accuracy	Macro F1
CNN	Piano-roll 88 × 512	0.984	0.984
LSTM	Piano-roll 88 × 512	0.830	0.832

Takeaway: with piano-roll inputs, the CNN's inductive bias (2D convs over time×pitch + residuals + SE) aligns better with musical texture and converges faster.
We use the CNN as the production model in the app.