SongSense: A Music Genre Classifier from Audio Signals

Sample post. This documents SongSense, a real project on this portfolio, written as a case study for the blog. Replace or expand it with more detail whenever you're ready — the full technical write-up is linked below as a PDF report.

01Overview

SongSense is a music genre classifier that takes an audio clip — uploaded or recorded live — and predicts its genre using audio signal processing and spectral analysis.

02Problem

Genre labels are useful metadata, but they're usually attached manually or inferred from surrounding text rather than the audio itself. A model that classifies genre directly from the signal doesn't depend on that metadata being present or accurate.

03Goals

• Classify genre directly from raw audio, not from external metadata
• Support both uploaded audio files and live recording as input
• Reach a reasonable classification accuracy across multiple genres

04Technologies

Python · Audio Signal Processing · Spectral Analysis

05Architecture

Incoming audio — whether uploaded or recorded live — is run through a spectral analysis stage that extracts frequency-domain features from the signal, which are then fed into a classification model trained to map those features to a genre label.

06Development

Most of the work here sits in the feature extraction stage rather than the classifier itself: getting the spectral representation of the audio right has a much bigger effect on classification quality than swapping in a fancier model on top of it.

07Results

88.4% classification accuracy across the target genre set.

08Lessons Learned

For audio classification specifically, the signal processing pipeline upstream of the model is where most of the leverage is — a clean spectral representation makes the classification step almost straightforward by comparison.

09Report

The full technical write-up is available as a PDF report.

10GitHub

The project repository is available on GitHub.