Automatic Raag Classification

1
Automatic Raag Classification
Parag Chordia, Alex RaeGeorgia Institute of
Technology
2
What is raag?
vadi
samvadi
Ascending C D G E F A G C Descending C B A G F
E D C
Scale
D G E F A G E F D C
Surface
G F E F
E F D C
E F A G
D G
G E
G F E F
Motives
3
Motivation

• Natural way of searching for ICM
• Highly correlated with mood descriptors
• Indian music auto DJ
• Starting point for most automated analyses

4
Problem

• Hard typically takes 5-10 years of active
  listening to discern large set of raags
• 100 common raags
• Shared gestures between raags
• Same notes used in different raags

5
Basic Idea

• Want a technique that does not rely on finding
  specific note sequences
• Note that use of phrases leads to a tonal
  hierarchy
• Pitch-class distribution has been effectively
  used to model tonal hierarchies in Western tonal
  music

6
PCD Feature Extraction
Divide into 30s frames
Pitch-track using HPS
Onset-detection (complex DF)
Histogram pitch values
7
PCD Example
8
PCDD Feature Extraction
Segment into notes
Assign single pitch-class to each note
CDGDGEFAGFGAGC
Construct bi-grams
Count bi-grams
9
Classification
PCD, PCDD feature vectors
10-Fold CV
Unseen
Split Criteria
Test data
Training data
SVM
MVN
FFNN
Random Forests
Trained model
Estimate error
10
Raag Database

• 31 different raags total of 20 hours
• 6 instruments
• sitar, sarod, bansuri, shenai
• vocal (male and female)
• Wide variety of musical forms
• different sections, densities, rhythms

11
(No Transcript)
12
Results
13
Discussion

• Nearly perfect in CV experiment
• PCDD increases performance by 20
• PCDD not generalizing well across recordings
• Why?
• sensitivity to onset detection errors
• However substantial melodic variation even in
  segments from same recording

14
Future Work

• Automatically identify tonic (sa)
• Robust to accompaniment
• Improve tonal onset detection
• Do we need more general n-gram model?