Columbia University

Technology Ventures

Cover song recognition software capable of searching vast music databases

Technology #cu12105

Questions about this technology? Ask a Technology Manager

Download Printable PDF

Collections
Image Gallery
Researchers
Daniel P.W. Ellis
Managed By
Satish Rao
Patent Protection
US Patent 9,384,272

Elucidating tracks in large music databases similar to a given song (specifically, cover versions) is a desirable process for applications such as copyright enforcement. Generally, this is an arduous and computationally expensive process particularly for large databases. For more extensive archive searching, a hashing approach exists wherein features extracted from a track can be compared to an index of corresponding features in the large database; however, current services using hashes can only match identical songs, not “similar” songs such as covers. This technology establishes a method to rapidly match similar versions of a given song, such as covers by various artists. This technology enables the matching of identical and similar versions of a particular song in an efficient manner capable of scanning large music databases.

Rapid algorithm adept for searching large song databases and matching similar versions of a given song

This technology rapidly generates chroma matrices comprised of 12-component vectors, which reflect a song’s musical content over time. From these chromas, the song recognition software generates jump codes based upon the amount of time elapsed between the song’s characteristic landmarks, such as fluctuations in pitch. These jump codes robustly discern changes in beat, harmony, and melody to accurately capture a song’s musical content, while remaining sufficiently general to allow comparisons with similar songs and high-throughput scanning of vast music database sets for similarity analysis.

The scalability of this song matching technology has been examined using the MillionsSong Dataset, which contains audio features and other kinds of metadata for over one million songs.

Lead Inventor:

Daniel P.W. Ellis, Ph.D..

Applications:

  • Searching large databases for similar songs
  • Musicology research
  • Recommending music to users
  • Identifying copyright infringements
  • Matching songs against those in video productions or from memory

Advantages:

  • Fast searching algorithm due to a sub-linear scaling approach as opposed to conventional linear algorithms.
  • Ability to search databases holding millions of songs

Patent information:

Patent Pending (US/2013/0091167)

Tech Ventures Reference: IR CU12105

Related Publications: