Source Codes for Chroma/Chord-based Large-Scale Audio Indexing/Hashing

Dear colleagues,

Source codes for Chroma/Chord-based Large-Scale Audio Indexing/Hashing are
released now, together with some example datasets of music songs. More
datasets can be provided under request. You can directly run programs by
following the instructions in "readme.txt" and can freely modify programs
to do any scientific experiments. The basic descriptions are listed below:

1. Indexing Based on Low-Level Music Features
Based on investigating the statistical representation of acoustic
short-time sequential correlations, a multi-probe histogram (MPH) is
computed from each audio track, to provide a more adequate balance between
scalability, robustness and discrimination ability. The idea of locality
sensitive hashing (LSH) was applied to compute MPH from a sequence of
chroma features. The major MPH bins of an audio track are in the top-n
major MPH bins of its variants with a high probability. Based on the
analysis of the order statistics (OS) of MPH bins, an adapted LSH approach
is suggested to map MPHs to hash values. (Download at
http://www.comp.nus.edu.sg/~yuy/MPH.zip [ www.comp.nus.edu.sg/~yuy/MPH.zip ] )

2. Indexing Based on Mid-Level Music Attributes
Chord progressions (CPs) are exploited to realize accurate and meaningful
summarization of music content and efficient organization of the
database.The SVMhmm model was adopted, SVM for per-feature chord
recognition, and HMM for CP recognition. Through a modified Viterbi
algorithm, N-best CPs are locally probed to generate a simple and
descriptive chord progression histogram (CPH). Organizing songs in the
layered tree-structure further helps alleviate the potential imbalance
among buckets.
(Download at http://www.comp.nus.edu.sg/~yuy/CPH.zip [ www.comp.nus.edu.sg/~yuy/CPH.zip ] )

Best regards,

Yi Yu