140 likes | 308 Views
Chord Recognition. EE6820 Speech and Audio Signal Processing and Recognition Mid-term Presentation JunHao Ip. Chord Recognition:. Introduction and Background Information Previous methods Feature description Experimental procedures. Introduction:.
E N D
Chord Recognition EE6820 Speech and Audio Signal Processing and Recognition Mid-term Presentation JunHao Ip
Chord Recognition: • Introduction and Background Information • Previous methods • Feature description • Experimental procedures
Introduction: • Music transcription >> Very difficult task >> Requires strong musical background and training • Chord transcription >> Some musical background >> Good ears >> Time consuming • Needed for automation >> Very challenging problem >> Limited successes • Investigate existing methods and try new techniques for Chord recognition
Musical Background: • Chord is a several tones played simultaneously. • It is usually played in a group of three tones called Triad. • Chord symbol is defined by root note and the key associated with it.
Previous Methods: • Manual transcription >> Currently the most accurate technique of all >> Strong musical training is needed >> Very time consuming • EM Trained Hidden Markov Model >> Compare performance of MFCC and Pitch Class Profile (PCP) >> PCP outperforms MFCC >> 83.3% accuracy in chord alignment, 26.4% accuracy in recognition • Chord Progression Hypothesis Model >> Attempt to make educational guess on chord progression >> Find keys, chord symbols, and beats concurrently >> Uses Chroma Vectors for chord estimation, similar to PCP >> 77% accuracy in recognition
Pitch Class Profile: • Combine pitches from different octaves to form a 12 semi-tones vector from (Ab to G#) • Feature is popular • Problem with data presentation when voice and multi instruments present. • Use 24 bins instead of 12 p [k] = floor( 24 log2( (k / N) (fs / fref) ) mod 24 PCP [p] = sum ( |X[k]|2 ) k = frequency index p = bin index Fs = sampling frequency Fref = 440Hz (note A)
Autocorrelation of Sub-band energy envelope: • New technique • Investigate correlation of sub-band energy envelope with long lag • The least common multiple of the different fundamental frequencies making up a chord Rxx[l] = ∑ x[n] x[n-l] = Rxx[-l]
Experiments: • Investigate the PCP feature • Sub-band Autocorrelation feature • Feature comparison • Train by Expectation Maximization Model? Nearest Neighbor? • Measure accuracy of Chord classification using test sets
Reference: Chord Segmentation and Recognition using EM – Trained Hidden Markov Model, Alex Sheh and Daniel P.W. Ellis, 2003 Automatic Chord Transcription with Concurrent Recognition of Chord Symbols and Boundaries, Takuya Yoskioka, Tetsuro Kitahara, Kazunori Komatani, Tetsuya Ogata, and Hiroshi G. Okuno, 2004 A Chorus-section Detecting Method for Musical Audio Signals Transcription Techniques – part 1, Lucas Pickfordhttp://www.globalbass.com/archives/dec2000/transcription_techniques.htm Introductory Musicianship A Workbook 6th Ed, Thomson Schirmer Inc, Theodore A. Lynn Speech and Audio Signal Processing, John Wiley & Sons Inc, Ben Gold and Nelson Morgan