Audio%20Fingerprinting

1
Audio Fingerprinting

• MUMT 611
• Philippe Zaborowski
• March 2005

2
What is Audio Fingerprinting?

• Every digital recording has unique features just
  like a human fingerprint.
• The audio fingerprint of a digital recording must
  first be extracted.
• Once extracted, it can be compared with reference
  fingerprints stored in a central database to
  identify the music.

3
Advantages of Fingerprinting

• Reduced storage requirements as fingerprints
  relatively small
• Efficient comparison as perceptual irrelevancies
  have been removed
• Efficient searching as the fingerprint database
  is also relatively small

4
Hash Functions vs. Fingerprinting

• In Cryptography hash functions allows comparison
  of two large digital files by comparing their
  hash values
• Very efficient way to determine whether or not a
  particular digital file is present in a large
  database

5
Hash Functions vs. Fingerprinting

• Hash functions cannot be used in Audio
  Fingerprinting
• Changing the audio format will change the digital
  waveform (CD to mp3 conversion)
• Changing a few bits would result in a completely
  different hash value

6
Hash Functions vs. Fingerprinting

• Fingerprinting does not establish mathematical
  equality, but perceptual equality
• Perceptually similar digital recordings will
  result in similar but not an identical
  fingerprint
• X and Y are similar F(X)-F(Y) lt T
• X and Y are not similar F(X)-F(Y) gt T

7
Fingerprinting Parameters

• Robustness
• Reliability
• Fingerprint size
• Granularity
• Search speed
• Scalability

8
Applications

• Broadcast monitoring
• Connected Audio
• Database maintenance
• Digital rights management
• Music library organization

9
Fingerprint Extraction

• Semantic features
• Include genre, beats-per-minute, mood
• Don't always have a clear meaning
• More difficult to compute
• Not universal (ex BPM in classical music)
• Non-semantic features
• Spectral flatness measure (Fraunhofer)
• Loudness, bandwidth (Bonn et al)
• Energy at each band (Haitsma/Kalker)

10
Fingerprint Extraction (Haitsma)

• Overlapping Hanning windows are used to extract
  32-bit sub-fingerprints every 11.6 ms
• Large overlap is used to smooth short varying
  differences in the waveforms
• FFT is used to find the band energy
• A fingerprint block consists of 256
  sub-fingerprints about 3 seconds in length

11
Fingerprint Extraction (Haitsma)
12
Fingerprint Extraction (Haitsma)
13
Database Search (Haitsma)

• Brute force method for matching
• 10,000 song database 250 million
  sub-fingerprints
• Searching through every fingerprint would take 20
  minutes!
• More efficient methods
• Look up table is 800,000 times faster!
• Requires 232 extra memory for the LUT
• Average of 300 comparisons needed

14
Database Search (Haitsma)
15
Commercial Products

• Phillips (Haitsma)
• Audible Magic
• Relatable (Music Brainz)
• Gracenote

16
Conclusion

• Many practical applications for both industry and
  consumers
• Very accurate and reliable, but not yet proven in
  real world
• May force file sharing underground