Impulse 2026 – Self-Supervised Audio Retrieval

This repository contains our submission for Impulse 2026, a national-level Signal Processing and Machine Learning hackathon focused on Audio Signal Processing.

Problem Overview

The objective is to learn meaningful audio representations without labels and use them for robust audio retrieval, similar to a Shazam-style identification system.

Methodology

We follow a structured, end-to-end design aligned with the problem statement:

• Phase 1 – Audio Representation

  • Raw audio preprocessing
  • MFCC-based time–frequency feature extraction
  • Data augmentations for self-supervised learning

• Phase 2 – Self-Supervised Learning

  • Contrastive learning framework
  • Neural encoder trained to bring augmented views of the same audio closer in embedding space

• Phase 3 – Audio Retrieval

  • Embedding database construction
  • Cosine similarity–based retrieval to identify the closest matching track

• Phase 4 – Embedding Analysis

  • PCA and t-SNE visualizations to study embedding structure

• Phase 5 – Qualitative Evaluation

  • Retrieval behavior analysis on representative samples

Repository Structure

• submission.py – Inference and retrieval script used for evaluation
• Impulse_2026.ipynb – Complete development notebook with experiments and visualizations
• outputs.csv – Test output file submitted for evaluation
• requirements.txt – Required Python dependencies

Dataset

We use the Free Music Archive (FMA) dataset for training and evaluation.
Due to size constraints, the dataset is not included in this repository.

Key Highlights

• Fully self-supervised (no labeled training data)
• Scalable retrieval pipeline
• Clear separation between training, analysis, and inference
• Reproducible and modular design

---