Advanced

Musical Instrument Categorization using Spectral- and Cepstral Analysis

Florén, Jakob and Torby, Joel (2016) MASY01 20161
Mathematical Statistics
Abstract
This thesis suggests and tests two methods for categorizing instruments in a musical signal. The
first method involves spectral analysis and the magnitude spectrum, analysing the structure of
pitches and overtones and how the amplitude of those changes over time. The second method involves
cepstrum analysis and especially the, for speech recognition popular method computing the
Mel Frequency Cepstral Coefficients (MFCC). For each method, a weighting procedure using Least
Squares is finally used to determine which instrument that are present in a signal and which are
not. This is done by comparing the unknown signal to a reference bank of tones. So, even if this
thesis is primarily about testing and comparing two methods, it is also... (More)
This thesis suggests and tests two methods for categorizing instruments in a musical signal. The
first method involves spectral analysis and the magnitude spectrum, analysing the structure of
pitches and overtones and how the amplitude of those changes over time. The second method involves
cepstrum analysis and especially the, for speech recognition popular method computing the
Mel Frequency Cepstral Coefficients (MFCC). For each method, a weighting procedure using Least
Squares is finally used to determine which instrument that are present in a signal and which are
not. This is done by comparing the unknown signal to a reference bank of tones. So, even if this
thesis is primarily about testing and comparing two methods, it is also about how much a linear
Least Squares weighting method can absorb the non-linear calculations of magnitude spectrum and
MFCC.
The results for both methods in combination with the Least Squares weighting procedure are
promising. For some musical signals the instrument categorisation succeeds very well while for
other signals improvements of the methods are needed. These improvements may be achieved by
combining features from both methods and adjusting the weighting system. (Less)
Please use this url to cite or link to this publication:
author
Florén, Jakob and Torby, Joel
supervisor
organization
course
MASY01 20161
year
type
M2 - Bachelor Degree
subject
language
English
id
8891409
date added to LUP
2016-09-12 13:35:47
date last changed
2016-09-12 13:35:47
@misc{8891409,
  abstract     = {This thesis suggests and tests two methods for categorizing instruments in a musical signal. The
first method involves spectral analysis and the magnitude spectrum, analysing the structure of
pitches and overtones and how the amplitude of those changes over time. The second method involves
cepstrum analysis and especially the, for speech recognition popular method computing the
Mel Frequency Cepstral Coefficients (MFCC). For each method, a weighting procedure using Least
Squares is finally used to determine which instrument that are present in a signal and which are
not. This is done by comparing the unknown signal to a reference bank of tones. So, even if this
thesis is primarily about testing and comparing two methods, it is also about how much a linear
Least Squares weighting method can absorb the non-linear calculations of magnitude spectrum and
MFCC.
The results for both methods in combination with the Least Squares weighting procedure are
promising. For some musical signals the instrument categorisation succeeds very well while for
other signals improvements of the methods are needed. These improvements may be achieved by
combining features from both methods and adjusting the weighting system.},
  author       = {Florén, Jakob and Torby, Joel},
  language     = {eng},
  note         = {Student Paper},
  title        = {Musical Instrument Categorization using Spectral- and Cepstral Analysis},
  year         = {2016},
}