General Research Interests:
My general interest is in machine
perception of music, that is to what extent is the information
contained in musical signals and musical scores accessible by
computer. My initial interest in this area was in pitch tracking by
computer both in the frequency domain and in the time domain and in
pitch perception by humans. Other problems I've studied have included
the use of autocorrelation to determine musical meter, pitch
perception of frequency modulated musical signals by skilled
performers, harmonicity of musical instruments, limits of accuracy in
performances and in pitch perception by skilled musicians, and the
effects of noise on accuracy of calculations using the phase vocoder.
Most recently I have been interested in musical instrument
identification by computer.
My initial results were an exploration of
the effectiveness of speaker identification techniques using pattern
recognition and a Gaussian mixture model with cepstral coefficients as
features. More recently I have examined the effectiveness of a number
of features in addition to cepstral coefficients, including
autocorrelation coefficients, spectral centroid, spectral
irregularity, and moments of the time wave, for the identification of
members of the woodwind family of musical instruments.
Publications are listed on my Wellesley page.
Recent Projects:
Musical instrument identification
a. Oboe/Sax identification
Calculations have been
carried out using cepstral coefficients and autocorrelation
coefficients as features to distinguish between sounds produced by
oboes and by saxophones. These computer calculations had a higher
success rate than human perception experiments on the same sounds.
b. Woodwind identification
I have examined the effectiveness
of a number of features in addition to cepstral coefficients,
including autocorrelation coefficients, constant Q coefficients,
spectral centroid, spectral irregularity, and moments of the time
wave, for the identification of members of the woodwind family of
musical instruments. Most effective was spectral irregularity (84 %
correct identifications) followed by cepstral coefficients and
autocorrelation coefficients.
Independent Component Analysis and Non-Negative Matrix
Factorization
With Paris Smaragdis we have successfully separated the notes of both trills and
more complicated polyphonic musical passages using ICA and NMF. We
can thus estimate the spectral profile and
the temporal information of every note in these examples.
Current Project:
Classification of Vocalizations of Killer Whales
A large number of whale sounds recorded from the Captive Killer Whale
Population at Marineland of Antibes, France were classified into call
types using acoustic input and visual examination of spectral patterns
[A. Hodgins-Davis, ``An Analysis of the Vocal Repertoire of the
Captive Killer Whale Population at Marineland of Antibes, France,''
thesis, Wellesley College, 2004]. The melodic contour of these
sounds was calculated using two methods of pitch tracking.
[J. C. Brown et al, "
Calculation of repetition rates of the
vocalizations of killer whales", JASA 116, 2004]. Using dynamic time
warping and a kmeans classifier, they have since been classified into
10 groups to compare with the 9 perceptual groups with roughly 90 %
accuracy. [J.C. Brown and P.J.O. Miller, "
Classification of
Vocalizations of Killer Whales Using Dynamic Time Warping", submitted
to WASPAA05] This method will be tested on other perceptually-grouped
killer whale sounds, and other means of classification eliminating the
tedious step of pitch tracking are being explored.