Fast, Robust Audio Fingerprinting: RARE

1
Fast, Robust Audio Fingerprinting RARE
Communication, Collaboration and Signal
Processing Group
C. Burges, J. Platt, J. Goldstein, E. Renshaw
Problem Statement
How Does It Work?
Each projection maximizes the signal-to-noise
ratio on a training set. A trace (64 floats),
computed every 186 ms, is compared to a database
of fingerprints.
Given a stream of possibly noisy audio, identify
known audio clips, and their position in the
stream
Some Applications
Performance

• Allow users to identify e.g. radio music with a
  PDA
• Allow music players to identify whatever is
  played, to provide e.g. artist, song name
• Automatically identify duplicates to clean a
  database
• Automate music detection to compute royalties
• Detect commercials to verify that they were
  aired, or to add metadata

• System can identify any one of 240,000 clips
  using 10 CPU on 850 MHz PIII
• Testing with 10 ½ days of audio (3,444 songs),
  using a confirmation fingerprint, gave a false
  positive rate of 1.5 x 10-8 per clip, per
  database entry, at a false negative rate of 0.2
  per clip. Testing on 36 hours of distorted audio
  gave error rates of 8 x 10-6 and 0.8
  respectively.

Compute Robust Projections
Database Group