Lecture Notes: Recreating Shazam's Song Recognition Algorithm

Introduction

• Motivation: Curiosity and desperation to land a junior Dev job.
• Shazam's tech fascination: Instantly recognizes songs using audio.
• Objective: Reverse engineer Shazam to demonstrate engineering skills.

Understanding Shazam's Technology

• Core concept: Audio Fingerprinting
  • Creates a unique 'DNA profile' for every song.
• Main Steps:
  1. Spectrogram Creation: Converts song to frequency content over time.
  2. Peak Identification: Pinpoints high-intensity frequencies or 'Peaks'.
  3. Hash Encoding: Encodes relationships into unique hashes linked to song metadata.
  4. Database Storage: Stores thousands of hashes as a song's fingerprint.

Process of Song Identification

• App records a snippet and creates its fingerprint.
• Uses fingerprints to query database for matches.
• Evaluates time coherence for candidate songs to find the correct match.

Implementation

• Language: Coded using Go, no libraries, for learning.
• Audio handling: Utilized libraries for heavy lifting tasks.

Key Functions in Algorithm

1. Spectrogram Creation:

   • Converts raw audio into spectrogram via sliding window technique.
   • Mitigates spectral leakage using Hamming window function.
   • Uses FFT to transform audio to frequency domain.
   • Downsamples audio, limits frequency range (20Hz to 5kHz).

2. Peak Extraction:

   • Applies filter of six logarithmic frequency bands.
   • Identifies loudest frequencies in each band.
   • Uses average of strongest frequencies as dynamic threshold.

3. Fingerprint Creation:

   • Treats each peak as an 'Anchor'. Identifies five nearby 'Targets'.
   • Creates unique hash encoding peak relationships.
   • Stores hashes in hashmap with time stamps and song IDs.

Front-End Development

• Used React for the user interface.
• Communication via WebSockets.
• Integrated Spotify for song data retrieval.
• Process:
  • Users upload songs or provide Spotify links.
  • Song data fetched from YouTube and processed.

Song Matching Process

• User records audio snippet.
• Snippet encoded and processed to create fingerprint.
• Queries fingerprint database for matches.
• Evaluates time coherence to find best match.

Conclusion

• Most challenging project undertaken.
• Provides a demonstration of technical capability.
• Links to demo and GitHub repository for further exploration.

Additional Notes

• Copyright laws prevent full demo.
• Demo video and code available via provided links.