Time Frequency Reassignment:

A Review and Analysis

Stephen Hainsworth Malcolm Macleod
CUED/F-INFENG/TR.459
.
1
Time Frequency Reassignment: A Review and Analysis
Stephen W. Hainsworth

and Malcolm D. Macleod

Cambridge University Engineering Department, Cambridge, CB2 1PZ, UK

Qinetiq, Malvern, WR14 3PS, UK

swh21@eng.cam.ac.uk,m.macleod@signal.qinetiq.com
Technical Report, Cambridge University Engineering Dept.
CUED/F-INFENG/TR.459
Abstract
Time-frequency reassignment is a relatively old but under-explored method for time frequency anal-
ysis. This report reviews previous research on reassignment and relates it to instantaneous frequency in
an explicit and novel way. New measures based on reassignment of the amplitude spectrum, as opposed
to the traditional phase spectrum, are proposed and analysed. The statistical properties of reassignment
as a sinusoidal estimator are compared with the Cramer-Rao bound and nally some applications of
reassignment in the eld of musical analysis are proposed.
1 Introduction
Many methods have been proposed for the task of sinusoidal estimation from noisy time series data.
Parametric methods such as AR estimation [50] or maximum likelihood [35] or model based methods
such as Bayesian MCMC methods [2] are one way of approaching the problem. These however, all suer
from model selection problems and/or are computationally very intensive. Non-parametric methods also
exist for the task and these are, in fact, more common. Many of these are based around the Fourier
transform, which is quickly implemented in the discrete domain via the fast Fourier transform (FFT).
However, the FFT leads to estimates of energy at xed frequency intervals and is hence limited in
resolution. To overcome this problem, many algorithms have been proposed (eg [34, 47]. This document
is concerned with one example of those: Time-Frequency Reassignment.
Time-frequency reassignment was rst introduced as a means of improving the readability of time-
frequency representations [33]. From the viewpoint of an engineer it may be considered to be a post-
processing step following the short-time Fourier transform (STFT). Whereas the STFT represents the
energy in a particular windowed signal as a set of point estimates on the time-frequency lattice (with
parameters determined by the block length and overlap), time-frequency reassignment shifts these coef-
cients away from the lattice to the centre of gravity of the windowed energy. Intuitively, time-frequency
reassignment uses information from the phase spectrum to sharpen the estimates in time and frequency
for each bin of the STFT. Unfortunately, this phase information, having been used to reassign the am-
plitude coecients, is no longer available for use in signal reconstruction.
The layout of this paper is as follows: section 2 will review the history behind reassignment and link
it to various instantaneous frequency measures. Then section 3 will present some novel measures which
correspond the reassignment of the amplitude spectrum. Section 4 reports some results on the statistical
analysis of reassignment as compared to the Cramer-Rao Bound and nally 5 describes some applications
of reassignment to musical audio analysis.

Work supported by the George and Lillian Schi Foundation

2
2 Review of Time-Frequency Reassignment
2.1 Historical Overview
The concept of time-frequency reassignment can be rst traced back to Kodera et al [33] in the 1970s.
Their focus was directed towards chirp-like signals and various methods of analysing the time-frequency
law associated with a signal were considered. This is not a unique denition of the signal but rather
an observation of it given an analysis technique and associated set of parameters. Obviously, dierent
methods and parameter sets will often (but not always) give dierent laws.
In [33], a number of methods for looking at simple signals were described and compared. The simplest
two methods were MS, the evolution of the maximum amplitude of the spectrum over time and ME, the
evolution in frequency of the maximum amplitude output from a bank of lters. The moving window
method (MWM) used a varying f for a two dimensional transform giving the distribution of energy
centred at t0 and f0 as follows:
s(t0, f0) =
_
f
0
+f/2
f
0
f/2
X(f) exp[2jft0]df (1)
The contour of the maximum was then followed using a variety of methods. In practice, the FFT was
used and the f then limited by the Heisenberg Uncertainty Principle.
The nal method proposed was to use the MWM method but to plot the energy at the centre of
gravity of the distribution

t,

f), not at the central point t0, f0).

t and

f are dened by

t = t0
1
2
arg|s(t0, f0)
f0
(2)

f =
1
2
arg|s(t0, f0)
t0
(3)
which are the group delay and instantaneous frequency respectively. Given these relocated time-frequency
points, a contour was again found and proved to be more concentrated and precise than for the MWM.
This was called the Modied Moving Window Method (MMWM) and the equations above encapsulate
the reassignment method. For various examples, Kodera et al showed that the MMWM method, utilising
reassignment, was the clearest method for displaying and analysing signals. The original paper [32] which
derived the MMWM method gives some implementation issues and shows that their original algorithm
worked on a dierence method, similar to those described below in section 2.2.3.
The next major step forward was many years later with several papers by Auger and Flandrin
[3, 4] where reassignment equations were derived for a number of dierent time-frequency and time-scale
distributions
1
. The treatment started with the Cohen class [14] of time-frequency representations
TFR(x; t, ) =
_ _
TF (u, )WV (x; t u, )du
d
2
(4)
where the distribution TFR(x; t, ) is considered to be the two-dimensional smoothing of the Wigner-
Ville distribution [49, 51]
WV (x; t, ) =
_
x(t + /2)x

(t /2)e
j
d (5)
by the kernel TF (u, ). This is desirable because the Wigner-Ville distribution has non-negligible cross-
terms which the smoothing kernel, if well chosen, can help to remove. Various kernels have been proposed
for dierent applications including the generalised WV distribution [6] originally applied to underwater
acoustic data and the modal transform [40] which was tailored specically for musical audio.
The disadvantage with smoothing is that it also broadens the signal components. This is because
even though the WV might not have any energy at a given point, there will likely be non-zero values
1
Also incidentally, they coined the term reassignment
3
nearby and these will have an inuence when the smoothing kernel is brought into play. Auger and
Flandrins suggestion was that instead of assigning the energy to the centre of the smoothing kernel, one
instead assigns it to the centre of gravity of these energy contributions:

t(x; t, ) = t
_ _
u.TF (u, )WV (x; t u, )du
d
2
_ _
TF (u, )WV (x; t u, )du
d
2
(6)
(x; t, ) =
_ _
.TF (u, )WV (x; t u, )du
d
2
_ _
TF (u, )WV (x; t u, )du
d
2
(7)
Figure 1 gives a toy example of this process.
Time
F
r
e
q
u
e
n
c
y
1 1.05 1.1 1.15 1.2 1.25 1.3 1.35 1.4
x 10
4
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Figure 1: A toy example of reassignment in action: the rectangle is intended to indicate the extent of the
smoothing kernel, the centre of which is marked by the cross. A traditional method would place the value
obtained from the convolution here. Reassignment would place it on the circle instead, where the centre
of gravity of the energy is located, this being oset both in time and frequency from the centre of the
convolution kernel.
The paper then went on to describe the application of these equations to various distributions, the only
one of which will be considered here being the spectrogram. This can also be considered to be a member
of the Cohen class of TF representations by using a smoothing kernel which is the WV distribution of a
suitable analysis window, h of unit energy:
TF (u, ) = WV (h; u, ) (8)
S
h
(x; t, ) =
_ _
WV (h; u, )WV (x; t u, )du
d
2
(9)
= [STFT
h
(x; t, )[
2
(10)
which is equivalent to the squared modulus of the short-time Fourier transform (STFT):
STFT
h
(x; t, ) =
_
x(u).h

(t u)e
ju
du (11)
4
0 200 400 600 800 1000 1200 1400 1600 1800 2000
0
0.5
1
0 200 400 600 800 1000 1200 1400 1600 1800 2000
2
1
0
1
2
x 10
3
0 200 400 600 800 1000 1200 1400 1600 1800 2000
400
200
0
200
400
sample no.
a) normal Hanning window, h
b) time derivative window, dh
c) time ramped window, th
Figure 2: Plots of the three windows used for reassignment. STFT
h
is generated using a) the normal
Hanning window; STFT
dh
is generated using b) the derivative window and STFT
th
is created using c) the
time ramped window.
In a further proof, they showed that the reassignment could be expressed purely in terms of spectro-
grams with dierent (but related) window functions:

t(x; t, ) = t !
_
STFT
th
(x; t, ).STFT

h
(x; t, )
[STFT
h
(x; t, )[
2
_
(12)
(x; t, ) = +
_
STFT
dh
(x; t, ).STFT

h
(x; t, )
[STFT
h
(x; t, )[
2
_
(13)
where STFT
h
is the standard short-time Fourier transform utilising a given window, STFT
th
is the STFT
using a time ramped version of the window and STFT
dh
is the STFT using the rst derivative of the
window function (which is equivalent to weighting the Fourier transform of the window function with a
frequency ramp). Figure 2 gives an example of these windows.
This formulation leads to an exact value of the phase derivative with no use of dierences as in
Kodera et al [32]. The above formulation does not require complex division but the equations can be
reformulated to the following slight simplication which does:

t(x; t, ) = t !
_
STFT
th
(x; t, )
STFT
h
(x; t, )
_
(14)
(x; t, ) = +
_
STFT
dh
(x; t, )
STFT
h
(x; t, )
_
(15)
Spectrogram reassignment, as expressed in the continuous case, gives perfect localisation for sinusoids,
impulses and chirps. It also retains the STFT properties of positivity and energy conservation but
loses bilinearity. It also dismisses the possibility of perfect reconstruction, which is probably the main
reason the method has been ignored for most applications (audio coding being an obvious example of an
application which requires this). Using reassignment on the spectrogram essentially incorporates phase
information into the display which is ignored in the traditional spectrogram. The phase of the STFT
contains valid information about the signal so this is obviously an improvement.
To illustrate the principles, several examples will be described: a stationary sinusoid would have high
localisation in frequency with bins being reassigned to the exact frequency but time reassignment would
be zero as the signal is symmetric in time. An amplitude modulated sinusoid would, however, exhibit
5
time reassignment; however, in the case where this modulation was symmetric around the centre of the
window, time reassignment would be again be zero. In the case of an impulse, frequency reassignment
would be zero, while time reassignment would be highly precise.
Reassignment was explored more completely by Chassande-Mottin [13, 9] under Auger and Flandrin.
He produced a geometric argument for the relationship between the reassignment vector and the 2-D
function of phase. Specically, in the case of a Gaussian window with unit variance, the reassignment
vector is simply the gradient of the scalar potential log[X[ where X(, t) is the STFT. For other windows,
there is an extra nonanalyticity factor.
Chassande-Mottin continued looking at reassignment and investigated the statistics of reassignment
vectors in certain cases [10, 12]. When just noise is considered, the vector is randomly distributed but
with a genuine signal present, this is not the case. Finally, this was developed into the concept of
supervised reassignment [11] where a number of dierent length STFTs were used and the statistics of
the reassignment vectors analysed to determine if there was a genuine signal present. Reassignment was
only carried out if this was the case, the aim of this being to prevent the generation of spurious artifacts
due to the reassignment process randomly clustering noise.
The above results were derived for continuous spectra, whereas in practice discrete signals and spectra
are more often encountered. All the reassignment equations can be applied using discrete signals, though
4 will show that small errors are encountered due to the sampling (dened explicitly by the Slepian-
Poldark-Landau theory [21]).
2.1.1 Fast approximation
Reassignment is already moderately ecient in terms of computational cost, compared to many algo-
rithms. However, in an unpublished study, de Rivaz [15] took the case of the Hanning window and
produced an approximation for the STFT
th
curve, which is close to being a scaled sine wave
STFT
th
(p)
3N
16j
(X(p 1) X(p + 1)) (16)
where X(p) is the unwindowed FFT value at index p, to give
(p) = (p) +

N
(S(p)) (17)

t(p) = t(p)
3N
8
!(S(p)) (18)
where
S(p) =
(X(p 1) X(p + 1)).STFT
h
(p)

2j[[STFT
h
(p)[[
2
(19)
This gives a very close approximation of the exact reassignment, as the frequency reassignment vector is
exact and only the time vector is slightly in error. As STFT
h
for the Hanning window can be calculated
easily from the unwindowed FFT as well, this formulation only actually requires one FFT calculation as
opposed to three.
2.1.2 Requantisation
Many versions of the reassignment algorithm to produce cleaned up spectrograms requantise the energy
from the new position to the closest STFT lattice point (eg [41, 5]). This is somewhat strange, because
it re-introduces smearing.
There are two possible justications for this requantisation. The rst concerns the Uncertainty
Principle, which states that the product of the support for a function in time and frequency has a lower
limit given by tf 1/4. The concepts surrounding this seemingly simple equation are surprisingly
complex. In the case of reassignment, the theory might be taken to state that the spectrogram grid,
as dened by the window support is the best one can do and therefore, the only meaningful positions
6
are those upon this grid. This is incorrect; a better statement is that there is an intrinsic and implied
quantisation grid given by the parameters of an undened, underlying Wigner-Ville spectrum. Each
point in the reassigned spectrogram will lie upon a grid point in an underlying WV spectra, but not
necessarily the same one. Thus, the reassigned spectrogram is not a valid time frequency representation
in the classical sense, because there is no underlying, all encompassing WV spectrum. By requantising
back to a spectrogram grid, all the energy is forced onto one coherent WV grid and the property is again
satised, though data is lost. It is worth noting that the lack of classical time-frequency validity in the
raw reassigned spectrum does not preclude its use as the rst stage of a later estimation system.
The other, more probable reason for requantisation is that it makes plotting functions much easier to
implement!
2.2 Instantaneous Freqency and how it relates to reassignment
Time-frequency reassignment can be interpreted as estimating the instantaneous frequency and group
delay for each point (bin) on the time-frequency plane. While group delay has been largely ignored,
various papers have explored the use of instantaneous frequency and its application to sinusoidal anal-
ysis. The purpose of this section is to relate some of these instantaneous frequency formulations to the
reassignment equations and show that they are equivalent.
2.2.1 Friedman
Friedman [22] in 1985 proposed using instantaneous frequency for speech analysis as method of sharpening
the spectrogram. The starting point of Friedmans formulation is the STFT dened as
X
h
(, t) =
_

x(t + )h()e
j
d (20)
which implicitly includes the moving of the window function h(t). He then considers the complex result
in polar coordinates with a time varying amplitude a(, t) and phase (, t)
X
h
(, t) a(, t)e
j(,t)
(21)
From this and using the denition that instantaneous frequency is the derivative of phase the following
can be proved:
(, t) =

t
(, t) (22)
=

t
|log X
h
(, t) (23)
=
_
1
X(, t)

t
X
h
(, t)
_
(24)
=
_
jX
h
(, t) + X
dh
(, t)
X
h
(, t)
_
(25)
= +
_
X
dh
(, t)
X
h
(, t)
_
(26)
This is the representation of frequency reassignment as derived by Auger and Flandrin several years later
for the case of the spectrogram (see equation 15).
2.2.2 Abe
Abe [1] in 1996 also produced a formulation for instantaneous frequency which instead utilised Cartesian
coordinates
X
h
(, t) = a + bj (27)
7
Following a similar method to Friedman,

t
X
h
(, t) =
a
t
+ j
b
t
jX
h
(, t) + X
dh
(, t) (28)
In Cartesian coordinates, the expansion on the instantaneous frequency is
(, t) =

t
arg X
h
(, t) (29)
=
a
b
t
b
a
t
a
2
+ b
2
(30)
as dened by Flanagan and Golden [20]
2
. This is where Abe stops but by combining equations 28 & 30,
the following steps lead back to the reassignment equation for frequency (conditioning on and t have
been dropped for convenience).
(, t) =
!|X
h
|jX
h
+ X
dh
|X
h
!|jX
h
+ X
dh

[X
h
[
2
(31)
=
!|X
h

2
+!|X
h
|X
dh
+ |X
h

2
|X
h
!|X
dh

[X
h
[
2
(32)
=
[X
h
[
2
+!|X
h
|X
dh
|X
h
!|X
dh

[X
h
[
2
(33)
= +
|X
dh
X

[X
h
[
2
(34)
2.2.3 Dierence methods
While the above formulations are both dierent, they can be seen to represent the same underlying
equation. They rely on the fact that X
dh
(, t), the STFT calculated with the derivative window, can be
exactly evaluated. In contrast, another set of papers such as Charpentier [8], Brown [7] and Dixon [16]
all use an approximation such that the dierential is replaced with a nite dierence.
(, t) =
(, t + t) (, t)
t
(35)
Dixon [16] uses brute force and calculates two STFTs centred one sample apart and nds the rate of
change of phase this way. Charpentier [8] and later Brown [7], on the other hand, exploit the property
of the Hanning window having a very compact analytic description in the frequency domain to perform
an extrapolation of phase at the next sample.
A survey paper on sinusoidal estimation by Keiler [30] compares a number of methods including
reassignment, an identical version to that described above in equation 35 and a version of Marchands
algorithm [36]. They speculate that these three are equivalent but are unable to prove this conjecture.
Further to the obvious relationship mentioned above between exact reassignment and the dierence
method,t is possible to relate Marchands formulation for sinusoidal estimation to frequency reassignment.
Marchands estimator is given by

f(k, n) =
fs

arcsin
_
1
2
[X
1
(k, n)[
[X(k, n)[
_
(36)
where X
1
(k, n) is the windowed FFT using a dierenced signal d(n) = x(n) x(n 1). It is trivial to
prove that X
1
(k, n) = X(k, n) X(k, n 1) and gure 3 shows the relationships between these complex
vectors, which form an isosceles triangle when the magnitude of the FFT vector is assumed to be constant
2
Though they used a dierence rather than a partial dierential.
8
over a window change of one sample. The phase change is given by 2 where is given by the arcsin
function above. Thus it can be seen that Marchands method is again using the rate of change of phase
in a dierence function to evaluate the instantaneous frequency. The method gives poorer results at high
frequency where the phase change between two samples approaches .
X
h
[k, n 1]
X
h
[k, n]

X
h
[k,n]X
h
[k,n1]
2

Figure 3: Figure showing vector relationships for Marchands method of nding instantaneous frequency.
While all these methods produce a measure of instantaneous frequency, it is only an approximation.
Equally ecient methods have already been described by the reassignment equations which nd the
exact IF, making the dierence methods somewhat redundant. Figure 4 shows the error between the
exact reassignment and Browns method. It was also discovered that Browns method was susceptible to
phase errors of 2.
The method of Charpentier can be linked with the work of Richard and Lengelle [45] who used a
similar recursion method. Their aim was to nd values of X(n + 1, ) where is h, dh and th so that
the reassignment vectors can be calculated at an advance of one sample, recursively. This was expanded
for various windows and for a number of dierent time-frequency distributions. However, the use of such
a ne hop rate between frames is of questionable benet.
2.3 Applications of Reassignment
As has already been mentioned, reassignment has been somewhat ignored over the years. Apart from
the various uses of instantaneous frequency for frequency analysis as mentioned in section 2.2, full reas-
signment has only been utilised since the mid 1990s.
Plante and Ainsworth [41] described time-frequency reassignment and then went on to use just fre-
quency reassignment to aid in formant analysis of speech. The interesting point to note about their work
is that they re-quantise the frequencies back to the STFT grid after reassignment. This was discussed
in section 2.1.2. A later paper by the same authors [42] gave examples of both time and frequency
reassignment used to sharpen up the spectrogram of speech.
Peeters and Rodet [39, 38] utilised both time and frequency reassignment for purposes of audio coding.
Frequency reassignment gave a measure of sinusoidality while time reassignment was used for nding local
energy maxima (for example, glottal pulses in speech).
In the context of music synthesis, Fitz et al [19] used reassignment on a spectrogram with signicant
window overlap to produce an accurate picture of the transient evolution of single notes. The analysed
9
0 100 200 300 400 500 600 700 800 900 1000
60
40
20
0
20
40
60
80
FFT bin number
e
r
r
o
r
,

H
z
Figure 4: Plot of error between Browns phase dierence method and the exact reassignment value. The
data was a single sinusoid at 400Hz with added white noise at 60dB; N = 2048.
results were then used for synthesis and perceptually better sounds were generated as a result.
Odegard et al [37] used frequency reassignment for seismic analysis and Goto [23] uses an instanta-
neous frequency method based on Abe [1] as a component in his program to extract melody and bass
fundamental frequencies from audio recordings. An application not involved with time series analysis
used reassignment to improve edge localisation in images [17]. Duvaut et al [18] also compared the re-
assignment method to Bayesian methods with application to gravitational waves, though they failed to
give a conclusion as to which is the best method.
3 Amplitude Reassignment
So far, we have introduced the two facets of reassignment, in time and in frequency. Frequency reas-
signment can be related to the familiar concept of instantaneous frequency (e.g. [22]), which is the
derivative of phase with respect to time as follows:
(, t) =

t
(, t) (37)
=

t
|log(X
h
(, t)) (38)
=
_
1
X
h
(, t)

t
X
h
(, t)
_
(39)
=
_
jX
h
(, t) + X
dh
(, t)
X
h
(, t)
_
(40)
= +
_
X
dh
(, t)
X
h
(, t)
_
(41)
10
where the STFT, X
h
(, t) is dened as
X
h
(, t) =
_
x(t + )h()e
j
d (42)
= e
jt
_
x()h( t)e
j
d (43)
leading to the derivation which allows eqn. 40 to follow from eqn. 39:

t
X
h
(, t) = jX
h
(, t)
+ e
jt
_
x()

t
h( t)e
j
d (44)
= jX
h
(, t) + X
dh
(, t) (45)
Similarly, time reassignment can be related to group delay of a signal (the derivative of phase w.r.t.
frequency) in the following manner:

t(, t) =

(, t) (46)
=

|log X
h
(, t) (47)
=
_
1
X
h
(, t)

X
h
(, t)
_
(48)
=
_
jtX
h
(, t) jX
th
(, t)
X
h
(, t)
_
(49)
= t !
_
X
th
(, t)
X
h
(, t)
_
(50)
where

X
h
(, t) =

e
jt
_
x()h(t )e
j
d (51)
= jte
jt
_
x()h(t )e
j
d
je
jt
_
x()h(t )e
j
d (52)
= jtX
h
(, t) jX
th
(, t) (53)
These formulations use the partial derivatives in time and frequency of the phase of the STFT.
Alternatively, one could consider the partial derivatives of the amplitude of the STFT:
t =

t
a(, t) (54)
= !
_

t
log X
h
(, t)
_
(55)
= !
_
jX
h
(, t) + X
dh
(, t)
X(, t)
_
(56)
= !
_
X
dh
(, t)
X
h
(, t)
_
(57)
11
=

a(, t) (58)
= !
_

log X
h
(, t)
_
(59)
= !
_
jtX
h
(, t) jX
th
(, t)
X(, t)
_
(60)
= !
_
jX
th
(, t)
X
h
(, t)
_
(61)
=
_
X
th
(, t)
X
h
(, t)
_
(62)
So what do equations 57 and 62 represent? It turns out that, after suitable scaling, they are actually
approximately equivalent to the traditional reassignment vectors; i.e. for frequency reassignment,

_
X
dh
(, t)
X
h
(, t)
_

_
2
Fs
_
2
N (63)
The measures t and are not in general, however, exactly equal to the traditional reassignment
vectors. An explanation for this can be constructed from the following observations: rstly, the amplitude
spectrum of the STFT is distorted by time-varying signal components in much the same way as the phase
spectrum (everyone has witnessed the spreading of the window function of time-varying signals). It can
be expected that this amplitude prole also encodes information about the characteristics of these signals
and an analysis of the bins surrounding a component signal will hopefully yield this information. Indeed,
the rate of change of amplitude in time for a given bin can intuitively be thought of as encoding a
measure of the group delay (eqns. 54 to 57) and the rate of change in frequency as relating to the
instantaneous frequency (eqns. 58 to 62). The second observation is that the noise statistics will combine
to give dierent characteristics in the phase and the amplitude spectra. This can account for some of
the dierence found.
Another reason for the dierences between these measures and the phase-derived reassignment vectors
can be seen by examining the Cauchy-Riemann equations. These state that for a function of a complex
variable fz(x, y) = u(x, y) + iv(x, y), where z = x + iy, if it is analytic then the derivative f

(z) is
independent of path used. Applied to the case in point, it means that
a(, t)
t

(, t)

(64)
a(, t)

(, t)
t
. (65)
The only analytic signal possible is the Gaussian. In all other cases there is what Chassande-Motin
[13] terms a non-analyticity factor which means that the Cauchy-Riemann conditions are not met.
A dierent perspective is that the unit variance Gaussian window has symmetric support in time and
frequency (i.e. the time support is equal to the frequency domain support). This symmetry extends to
the phase space and amplitude space which become functions of each other and the phase and amplitude
reassignment vectors are identical
3
. This can be compared with the uncertainty principle where Gaussians
are the only type of signal which meet the equality; all others have a covariance term which increases
the uncertainty.
Despite these dierences, the measures in eqns. 57 and 62 can be used as vectors for time and
frequency reassignment respectively. From here on, the result of these equations shall be referred to as
the amplitude reassignment measures, as opposed to the traditional phase reassignment ones, because
they nd values for the group delay and instantaneous frequency via reassigning the amplitude spectrum.
3
This is true for the continuous Fourier transform but for the DFT, truncation eects will mean that it is again only
approximate.
12
Figure 5 shows plots of frequency reassignment using both amplitude and phase vectors. It can be seen
that the reassigned spectra are fairly similar, though there are dierences as expected, often at low
signal to noise ratio regions of the spectrum. This leads to another observation: it would be possible to
robustify a method for sinusoidal detection by using both phase and amplitude reassignment measures
to search for regions where they both have similar values. This could be applied to the work in [28] on
sinusoid/transient/noise classication.
0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000
10
4
10
2
10
0
10
2
Amplitude reassignment
0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000
10
4
10
2
10
0
10
2
Phase reassignment
frequency, Hz
Figure 5: Plot showing frequency reassignment using a) amplitude and b) phase
4 Reassignment and CRBs
4.1 Introduction
Section 2 examined the derivation and applications for reassignment, but the question of how well it
performs as an estimator has been left unanswered. In the literature, this problem has been ignored,
with the success of reassignment algorithms being judged by graphical means. This is not a quantitative
method so in this section the performance of reassignment against the statistical test of the Cramer-Rao
Bound (CRB) will be examined. Further work on this topic can be found in [26].
CRBs are a means of establishing the best possible performance of an estimator algorithm, given a
certain data set. They are dened in terms of unbiased estimation error variance (though estimation bias
is also an important eect which will be discussed below). Typically, the algorithm is then tested in a
series of trials for comparison with the CRB and to other methods performing on the same data set. In
the case of sinusoidal parameter estimation, much of the early work was done by Rife and Boorstyn [46]
where they proposed a signal model for real sinusoids
s(t) =
k

i=1
bi cos(it + i) (66)
13
where bi, i and i are the amplitude, frequency and phase of the sinusoid. For complex sinusoids, the
following was used
c(t) = s(t) + j s(t) =
k

i=1
bi exp[j(it + i)] (67)
where s(t) is the Hilbert transform of the real signal; c(t) is often said to be the analytic signal. From
this, Rife and Boorstyn went on to derive simple expressions for the CRBs of maximum likelihood (ML)
estimation on the discrete sampled equivalent of these equations. ML is the best minimum variance
unbiased estimation one can do given a data set
4
and they examine an almost-ML algorithm and its
performance with various windows.
The CRB measure of performance is estimation variance but bias is equally equally important. Bias
can be thought of as the mean estimation error and variance is then a measure of the trial to trial
variation. With closely spaced multiple tones, bias is the dominating eect. The use of windows reduces
the bias but increases the variance - the CRBs as derived by Rife and Boorstyn are for unwindowed
(or rectangular windowed) data and this gives the worst bias. When detection is an issue as well as
estimation, the trade o these eects has to be considered and windows with the best characteristics for
the given task are often an important consideration.
Real signals often contain multiple tones which are closely spaced (especially music), non-stationary
components and often have coloured noise to add to the diculties of detection and estimation.
Since Rife and Boorstyn, various estimators have been proposed and their performance related to the
CRB, including Jain [29], Grandke [24] (who concentrated on the Hanning window), Quinn [43, 44] (who
used phase as well as amplitude interpolation) and Macleod [34].
4.2 CRBs and Reassignment for a single complex tone
The simplest case to consider is the parameter estimation of a single complex tone under dierent noise
conditions
5
. The well-known ML estimate of the signal
c[n] = b1 exp[j(1n + 1)] + v[n] (68)
where v[n] is i.i.d. additive Gaussian noise of variance
2
, is the one which minimises
N1

n=0
v
2
[n] =
N1

n=0
_
c[n]

b1 exp[j( 1n +

1)]
_
2
(69)
This is often achieved via a coarse search of the periodogram followed by a ne search to estimate 1
accurately [34]. The CRBs are found as follows: if the parameters to be estimated are formed into the
vector
= [ 1 b1 1 ]
T
(70)
then the Fisher information matrix, J is dened as having elements
Jij =
1

2
N1

n=0
!
_
e[n]
i
e

[n]
j
_
(71)
where e[n] = b1 exp[j(1n + 1)]. From this, the CRB for a given parameter is given by
var| i (J
1
)ii (72)
4
This is assuming that no prior knowledge is used, as in, for example, Bayesian MCMC [2].
5
A real tone consists of two complex tones which have a bias eect on each other.
14
i.e. the i
th
diagonal element of the inverse of J. The bound on frequency estimation for a single tone
can be expressed as

1

6
2
b
2
1
N(N
2
1)
(73)
This can be seen to be inversely proportional to the signal to noise ratio (SNR) [34].
SNR =
b
2
1

2
(74)
A plot of this bound verses SNR can be seen in g. 9.
Reassignment performs two functions: it estimates instantaneous frequency and group delay. There
is no published CRB formulation for the latter so we concentrate on the former. Reassignment also does
not estimate amplitude nor phase so we shall ignore the bounds on these estimates as they are identical
to those achieved by the windowed FFT. Thus, the interesting bound is that on frequency estimation as
given by equation 73.
The rst interesting discovery is that, when the reassignment method is used, it turns out to have a
bias eect on frequency estimation and this estimate is dependent upon the choice of window used. This
seems to be in direct contradiction to the statement made by Chassande-Mottin et al [11] ...spectrogram
reassignment of a chirp can be proved not to depend on the used window... and also the perfect local-
isation property stated by Auger and Flandrin. This eect is not predicted in the theory because this
was derived by Auger and Flandrin [4] in the continuous domain. The bias is introduced as an artifact
when a discrete formulation is used.
The eect can be explored as follows: without loss of generality, we shall consider a complex sinusoid
with unit amplitude and phase equal to zero, e1[n] = exp(j1n). The DTFT is then

E1() =
N1

n=0
exp(j1n) exp(jn) (75)
=
_
exp(jN
d
/2)
exp(j
d
/2)
__
sin(jN
d
/2)
sin(j
d
/2)
_
(76)
where
d
= 1 . If we now consider the DFT, as calculated by the FFT, E1[k] =

E1(k0), where k0
is the bin frequency. Following this, if we consider 1 to be the combination of an exact bin frequency
plus an oset, it can be expressed as 1 = (p +)0 where is the oset from the exact bin. The DFT
of a general bin, k = p + m can then be expressed as
E1[p + m] =
exp(j) sin()
exp(j(m)/N) sin((m)/N)
(77)
Finally, we can now use these to complete the formulation for reassignment, in this case using the Hanning
window.

r =
N
2

_
X
dh
X
h
_
(78)
=
N
2

_
j
N
(E1[p + 1] E1[p 1])
E1[p]
1
2
(E1[p + 1] + E1[p 1])
_
(79)
=
N
2

_
2j
N
sin
_

N
_
sin
_
2
N
_
exp
_
j
N
_
sin
2
_

N
_ _
1 + exp
_
2j
N
__
_
. (80)
If rst order Tailor expansions are made for each of the terms in (80), the approximation,

r
N
2

_
2j
N

N
2
N
_
1 +
j
N
_

2
N
2
_
1 + 1 +
2j
N
_
_
(81)
= (82)
15
window bias (in units of bins)
Hanning 0.00007531015773
Hamming 0.05999774838755
Blackman 0.14265043248416
Table 1: Comparison of relative bias magnitude for dierent windows with F
s
= 4kHz and N = 256 samples.
0.5 0.4 0.3 0.2 0.1 0 0.1 0.2 0.3 0.4 0.5
2.5
2
1.5
1
0.5
0
0.5
1
1.5
2
2.5
, bins
b
i
a
s
,
H
z
Hamming
Blackman
Hanning
Figure 6: Shape of bias curves for the Hanning, Hamming and Blackman windows. The Hanning window
has a similar shape to the Hamming window. f
s
= 4kHz, N = 256.
can be derived. However, the higher order terms which have been ignored in eqn. 81 combine to give
the bias observed in reassignment.
Various windows have been investigated and normalised plots of the bias shape for several examples
can be seen in g. 6. Thus it can be seen that for the Hanning and Hamming windows, the bias has
the eect of slightly underestimating frequency, whereas for the Blackman window, the estimation of
delta is actually over the exact frequency. Table 1 gives the relative magnitudes of bias for the three
considered windows and is also inversely proportional to N, the number of samples in the block. This
is demonstrated in g 7 for the Hamming window. Thus it can be seen that if a moderately long block
length (N = 1024 samples) the bias is actually very small (of the order of 1/60th of a bin even for the
Blackman window).
Various windows were investigated and plots of the bias shape are displayed in gure 6. For the
Hanning and Hamming windows, the bias has the eect of slightly underestimating the magnitude of ,
whereas for the Blackman window, [[ is actually over-estimated. The relative intrinsic bias magnitudes
for the Hanning, Hamming and Blackman windows are approximately 1:800:1900. Bias is also inversely
proportional to N, the number of samples in the block; this is demonstrated in gure 7 for the Hamming
window. Thus it can be seen that if a moderately long block length (N = 1024 samples) the bias is
actually very small (of the order of 1/60th of a bin even for the Blackman window).
Comparison to the CRB without the removal of this bias is meaningless as at high SNR it will become
the dominating eect. Thus, a scheme to remove it is proposed as follows:
1. In the case of no noise, nd the bias needed to correct for each estimated value of to the correct
value.
2. Parameterise the resulting curve as a polynomial of suitable order (determined empirically).
3. When estimating the frequency, the evaluate the polynomial using the estimate of to nd the
bias, which is then corrected for.
After removal of the bias the performance of the reassignment method can be compared meaningfully
16
0.5 0.4 0.3 0.2 0.1 0 0.1 0.2 0.3 0.4 0.5
2
1.5
1
0.5
0
0.5
1
1.5
2
, bins
N = 1024
N = 512
N = 128
N = 256
b
i
a
s
,
H
z
Figure 7: Bias verses delta for dierent Hamming window lengths, N (f
s
= 4kHz).
20 0 20 40 60 80 100
10
4
10
2
10
0
10
2
10
4
SNR, dB
v
a
r
i
a
n
c
e
,

b
i
n
s
Figure 8: CRB comparison for a single complex tone using reassignment with dierent windows. (a) - CRB
of ML estimator; (b) - -, Hanning window; (c) -, Hamming window; (d) -, Blackman window.
17
10 0 10 20 30 40 50 60 70 80
10
4
10
3
10
2
10
1
10
0
10
1
10
2
SNR, dB
v
a
r
i
a
n
c
e
,

b
i
n
s
37.8 38 38.2 38.4 38.6 38.8 39
10
1.7
10
1.6
10
1.5
SNR, dB
v
a
r
ia
n
c
e
, b
in
s
Figure 9: CRB comparison for a single complex tone. (a) - CRB of ML estimator; (b) x, Marchands
estimator; (c) , reassignment using Hanning window; (d) , dierence method; (e) o, Grandke interpolator;
(f) +, Macleod 3-sample interpolator.
with the CRB. Figure 8 shows the results using reassignment with dierent windows assuming correct
detection and with bias removed. It can be seen that at high and low SNR, the comparatively high
bias found for the Hamming and Blackman windows cannot be eectively removed completely
6
. The
other interesting observation which is not evident from the plot is that in mid range SNR, the Hamming
window actually slightly outperforms the Hanning window by about 0.15dB. The Blackman window is
only marginally behind the Hanning window in performance.
Next, performance of Hanning reassignment is compared to the other estimators mentioned above.
Figure 9 shows the results of the various estimators as they break down due to noise. As the SNR
increases, an asymptotic tendency with respect to the CRB
7
is found, with an oset which is due to the
particular method used (related to the windowing). The results here are independent of window length
and sample frequency used though they were generated using N = 128 samples and fs = 4kHz. They
were also averaged over a range of signal frequencies and inter-bin osets , though again, the estimator
performance was independent of these parameters. The three methods which use phase all perform
almost equivalently, which is not surprising, given the underlying similarity, though reassignment is
marginally the best. They all achieve slightly better performance than the Grandke interpolator by
0.15dB. As expected, the Macleod 3-sample interpolator using unwindowed data gets signicantly closer
to the CRB.
Figure 10 shows the performance of each estimator when detection is taken into consideration. For
this purpose, the simplest detection scheme of picking the highest peak was used. The threshold at
-1dB SNR can clearly be seen, at which point the probability of detecting the wrong peak becomes
signicant. As the SNR falls, the variance rises rapidly, eventually achieving the second asymptote which
is equivalent to picking an arbitrary frequency. The performance of Marchands algorithm in this failure
region is articial: numerical issues with the algorithm mean that it is impossible to pick a value around
the Nyquist frequency so the results are being averaged over a smaller range.
6
For the Hanning window, even at 200dB SNR, the bias is completely removed.
7
All plots show the CRB for unwindowed data, though is possible to calculate the CRB for windowed data specically.
18
15 10 5 0 5 10 15
10
1
10
0
10
1
10
2
10
3
SNR, dB
v
a
r
i
a
n
c
e
,

b
i
n
s
Figure 10: CRB comparison for a single complex tone with detection. (a) - CRB of ML estimator; (b) x,
Marchands estimator; (c) , reassignment using Hanning window; (d) , dierence method; (e) o, Grandke
interpolator; (f) +, Macleod 3-sample interpolator.
4.3 Two complex tones
With the addition of a second complex sinusoid, the problem complexity increases dramatically. Before,
the bias for each tone depended solely upon , the bin oset for the single true tone. With two tones there
are also traditional bias eects coming into play, which are inherent in any nite windowed Fourier-based
method. The bias is now dependent upon the following:
f, the frequency separation of the two tones
i, the bin oset of the tones
r, the relative phase of the two tones
ar the relative amplitude of the two tones
The frequency separation has an eect rather like that described by Rife and Boorstyn [46], which
is due to the interference of the spectral window tails of the two sinusoids. This is roughly inversely
proportional to separation, though the function has maxima and minima which are related to the
f
,
the oset from an exact number of bins of this frequency separation. This is the dominant bias eect.
The bin oset has an eect rather like in the single tone case, in that it modulates the bias due to
the frequency separation. The relative phase and amplitude of the two tones have a much more complex
eect on the bias.
There are various observable variables which can be used to attempt bias removal. Frequency sep-
aration of any two bins is one observable, as are the i values of those bins. These will not, however,
correspond to the underlying f and is. A relative amplitude can be estimated, though the relative
phase is very hard to nd. There is also the issue of which bins to use in the estimation: with two
clear peaks, it makes sense to use these maxima, but often the phase relationship combines to produce
a central bin which forms a single peak.
The overall conclusion, after some investigation is that there is no simple analytical expression for
the bias, because there are too many parameters upon which it depends. A potential method of solving
19
the problem would be to use lookup tables for a large number of relative phases and amplitudes and
perform interpolation between these but this is likely to be successful as there are multiple combinations
of underlying parameters which will give similar observable variables.
4.3.1 Bias removal
As an example, a single case was investigated: to allow a comparison with the statistics of Rife and
Boorstyn [46], the case of equal amplitude and worst case phase was chosen. The method of bias removal
used was then to produce a family of surfaces which can be used as a look up table for the bias. These
were again produced in the case of zero noise. Bias removal proceeds as follows:
1. For a sensible pair of bins, nd

f and

i.
2. Use these to iterate around a pair of surfaces which give a better estimate of

f and

i.
3. Use these to index a third surface to nd the estimated bias.
This is a very quick summary of what is actually a far more complex procedure. For instance, the pair
of bins chosen in step 1 leads to one of three surface sets which are used; the upper frequency cisoid
is indexed by to nd its bias; the surfaces are not completely continuous and this leads to some
occasional errors.
The net results of this algorithm can be seen in g. 11. The block length was 128 samples, Fs =
4000Hz, and 6000 trials were made over a range of for each separation, f and at each value of SNR.
It is obvious that the results are above the CRB, which is the lower surface, at all points, except at
separations of less than 20Hz where the algorithm seems to better it. This is due to the fact that the
algorithm is not performing detection; it is making a two tone assumption. If this presumption was not
made, then it is probable that a detection algorithm would nd only one sinusoid. Many estimation
algorithms do not present results for separations this small.
The other eect that can be observed is that at some separations the experimental results do not
improve at the same rate as the CRB. This is mainly due to mis-estimations at the extremes of
(choosing = -0.5 rather than +0.5 and vice versa) on some of the trials.
The expansion of this method to general multiple tone bias removal is, as has already been mentioned,
non-trivial and has not been explored.
4.4 Multiple tone estimation - pseudo-real signals
Many signals, including musical ones which are the overall focus of the research in this project, consist
of multiple tones, often randomly spaced, though sometimes having a regular harmonic structure. To
test the performance of reassignment on more realistic examples a piano tone from the McGill Master
Samples collection [48] was analysed and the amplitudes of the rst twenty one harmonics manually
extracted. These were used along with a noise oor which was 70dB below the maximum amplitude
(matched to the original spectrum by hand) to construct a synthesised example with a known ground
truth - real-world signals have the disadvantage of unknown exact parameters.
For a range of fundamental frequencies from 200Hz to 400Hz trials were performed with random
phase relationships between the harmonics. The average of these trials is plotted in g. 12, against the
CRB for single tone estimation, given the theoretical SNR of each sinusoid from the noise oor. The
equivalent estimation performance of Macleods Hanning window estimator [34] is also included.
This plot does not separate bias and variance eects, but rather plots them together. On further
analysis, the bias and variance turn out to contribute about equal parts to the overall error for the
reassignment estimator. The Macleod algorithm out-performs reassignment at high SNR but gives about
equal performance at low SNR; this is because the Macleod algorithm performs iterative analysis in order
to remove bias.
20
0
5
10
15
20
25
30
35
0
100
200
300
4
3
2
1
0
1
2
3
4
5
SNR dB
fr
e
q
u
e
n
c
y
d
iffe
r
e
n
c
e
, H
z
v
a
r
i
a
n
c
e
,

H
z
Figure 11: Results of the two tone reassignment frequency estimation algorithm. Axes are SNR, f and
estimate variance in Hz. The CRB is plotted for comparison as the lower, blue surface. Parameters of the
tones are equal amplitude and worst case phase. F
s
= 4kHz and N = 128 samples.
21
0 5 10 15 20 25
10
5
10
4
10
3
10
2
10
1
10
0
10
1
Harmonic number
v
a
r
i
a
n
c
e

(
H
z
)
Figure 12: Simulation results for a synthesised piano tone. Black - CRB; Red .- Macleod algorithm; Blue
reassignment; Yellow - reassignment using zero padding.
22
4.5 CRB analysis conclusion
In conclusion for the Cramer-Rao Bound analysis, for single tones, reassignment out-performs other
analysis methods, after the removal of a bias which is inherent in the algorithm. However, for multiple
tones, the between-tone bias complicates estimation to the point where bias removal becomes almost
intractable. In a pseudo-real example, reassignment also fails to perform at an equivalent standard to
the algorithm of Macleod [34]. However, the inclusion of iterative tone removal and subsequent re-
analysis as used by Macleod may well explain this additional performance. If this was also used with
reassignment, it could improve the performance signicantly, though this is left for future work.
8
5 Applications
So, what use is time-frequency reassignment? A recent survey on sinusoidal parameter estimation by
Keiler [30] found that reassignment performed generally better than a number of other algorithms in-
cluding Keilers own triangle algorithm [31] and a parabolic interpolation method. This is in agree-
ment with the CRB analysis above and it seems that to attain additional performance, an iterative
estimation-subtraction-reestimation method has to be used. Therefore reassignment is potentially useful
as a sinusoidal estimator which is computationally ecient.
However, this does not take account of the facet of reassignment which has been ignored for some
time in this paper - that of time reassignment. Here, the method intrinsically oers another level of
improvement over the standard STFT which is completely unavailable to other methods. Indeed, time
reassignment is possibly more useful than frequency reassignment.
Musical audio is a challenging signal to analyse and reassignment oers some uses in this area. An
earlier paper by the authors [28] describes a method for classifying regions of the spectrogram to be
one of sinusoid, transient or noise by a simple application of a linear Gaussian decision boundary in
a multidimensional feature space. Figure 13 shows the output of this algorithm for a slap bass note
where a measure of component variance (or sinusoidality) extracted from the reassigned parameters is
also plotted; the harmonics can be seen descending towards the noise oor and becoming less sinusoidal
and also an almost undetectable transient (to the human ear at least) has been found at around 7s.
Another use of reassignment is in transient location for beat tracking where many methods make rm
detection decisions and attempt to track these onsets through time. The tracking algorithms can be
highly sensitive to timing errors and time reassignment can be used to nd a sub-frame estimate for the
onset. Another similar use is in nding a continuous measure for change detection: at frames of high
signal variability, the mean time reassignment is large and this measure can be used for change detection
[25]. Figure 14 gives an example of this. Possibly counter-intuitively, this measure works better with
longer frame lengths.
Also, the measure of time reassignment derived from the amplitude spectrum (t from eqn. 54) has
been found to be a good measure of sinusoidality and can be used for rough detection and removal of
harmonic information, which again has application in nding onsets. This has been described more fully
in [27].
6 Conclusions
In conclusion, this report has reviewed the history of the time-frequency reassignment method and related
it to a large body of work on instantaneous frequency and also a number phase dierence methods
which turn out to be sub-optimal versions of frequency reassignment. The next section presented some
new theory which shows that the amplitude spectrum can also be utilised to provide measures akin to
traditional phase reassignment and following this, Cramer-Rao analysis of frequency reassignment was
detailed. Finally, some uses of reassignment in musical audio were briey described, mainly utilising time
8
A temporary note is that [26] contains a fuller discussion of these issues and my thesis will do so even more, both coming
to dierent conclusions.
23
0 1 2 3 4 5 6 7 8
0
200
400
600
800
1000
1200
1400
1600
1800
2000
2200
Output of program
time, s
f
r
e
q
u
e
n
c
y
,

H
z
Figure 13: Slap bass note output with component (vertical) width proportional to frequency variance
reassignment which is the facet which is obviously superior to most existing FFT based methods. Thus,
reassignment can be concluded to be a useful tool for signal analysis, coming some way back to regaining
the precision of a Wigner-Ville transform without the cross terms and at a much cheaper computational
cost.
24
Time
F
r
e
q
u
e
n
c
y
0 1 2 3 4 5 6 7 8 9 10
0
0.5
1
1.5
2
x 10
4
0 50 100 150 200 250 300 350 400
200
100
0
100
200
300
400
500
time (frame number)
Figure 14: Spectrogram of signal and the time reassignment measure of signal change.
1
References
[1] T. Abe, T. Kobayashi, and S. Imai. Robust pitch estimation with harmonics enhancement in noisy
environments based on instantaneous frequency. In Proc. ISCLP, pages 127780, 1996.
[2] C. Andrieu and A. Doucet. Joint Bayesian model selection and estimation of noisy sinusoids via
reversible jump MCMC. IEEE Trans. Signal Processing, 47(10):266776, October 1999.
[3] F. Auger and P. Flandrin. Generalization of the reassignment method to all bilinear time-frequency
and time-scale representations. In Proc. ICASSP, volume IV, pages 31720, 1994.
[4] F. Auger and P. Flandrin. Improving the readability of time-frequency and time-scale representations
by the method of reassigment. IEEE Trans. Signal Processing, 43(5):106889, May 1995.
[5] F. Auger, P. Flandrin, O. Lemoine, and P. Gonalvs. Time-frequency toolbox for matlab. available
on-line at http://crttsn.univ-nantes.fr/ auger/tftbftp.html.
[6] B. Boashash and G. Frazer. Time-varying higher-order spectra, generalised Wigner-Ville distribution
and the analysis of underwater acoustic data. In Proc. ICASSP, volume V, pages 1936, 1992.
25
[7] J.C. Brown and M.S. Puckette. A high resolution fundamental frequency determination based on
phase change of the Fourier transform. J. Acoust. Soc. Am., 94(2):6627, August 1993.
[8] F.J. Charpentier. Pitch detection using the short-term phase spectrum. In Proc. ICASSP, pages
1136, 1986.
[9] E. Chassande-Mottin. Methodes de reallocation dans le plan temps-frequence pour lanalyse et le
traitment de signaux non stationnaires. PhD thesis, Universite de Cergy-Pontoise, 1998.
[10] E. Chassande-Mottin, F. Auger, and P. Flandrin. Statistique des vecteurs de reallocation du spec-
trogramme. Technical Report 96-01, Laboratoir de Physique, ENS Lyon, 1996. jul.
[11] E. Chassande-Mottin, F. Auger, and P. Flandrin. Supervised time-frequency reassignment. In IEEE
Int. Symp. on Time-Frequency and Time-Scale Analysis, pages 51720, Paris, 1996.
[12] E. Chassande-Mottin, F. Auger, and P. Flandrin. On the statistics of spectrogram reassignment.
Multidimensional Systems and Signal Processing, 9:35562, 1998.
[13] E. Chassande-Mottin, I. Daubechies, F. Auger, and P. Flandrin. Dierential reassignment. IEEE
Signal Processing Letters, 4(10):2934, October 1997.
[14] L. Cohen. Generalised phase-space distribution functions. J. Math. Phys., 7:7816, 1966.
[15] P. de Rivaz. Fast fourier reassignment. Unpublished; personal communication.
[16] S. Dixon. Extraction of musical performance parameters from audio data. In Proc. IEEE Pacic-Rim
Conf. on Multimedia, pages 4245, Sydney, Australia, December 2000.
[17] T. Dorney, S. Bhashyam, A. Doran, H. Choi, P. Flandrin, and R. Barnaniuk. Edge localized image
sharpening via reassignment with application to computer tomography. In Proc SPIE 45th Annual
Meeting, pages 41212, San Diego, CA, 2000.
[18] P. Duvaut, A. Doucet, C. Veaux, and P. Flandrin. Instantaneous frequency estimation: Bayesian
approaches verses reassignment - application to gravitational waves. In Proc. ICASSP, volume 5,
pages 296871, 1996.
[19] K. Fitz, L. Haken, and P. Christensen. Transient preservation under transformation in an additive
sound model. In Proc. International Computer Music Conference, 2000.
[20] J.L. Flanagan and R.M. Golden. Phase vocoder. Bell System Technical Journal, 45:14931509,
November 1962.
[21] P. Flandrin. Time-Frequency/Time-Scale Analysis. Wavelet Analysis and Its Applications. Academic
Press, 1999. Translated from the French (1993) by Joachim Stockler.
[22] D.H. Friedman. Instantaneous-frequency distribution vs time: an interpretation of the phase struc-
ture of speech. In Proc. ICASSP, pages 11214, 1985.
[23] M. Goto and S. Hayamizu. A real-time music scene description system: detecting melody and bass
lines in audio signals. In Proc. IJCAI Workshop on CASA, pages 3140, 1999.
[24] T. Grandke. Interpolation algorithms for discrete Fourier transforms of weighted signals. IEEE
Trans. Instr. and Meas., IM-32(2):3505, June 1983.
[25] S.W. Hainsworth and M.D. Macleod. Automatic bass line transcription from polyphonic music. In
Proc. International Computer Music Conference, Havana, Cuba, 2001.
[26] S.W. Hainsworth and M.D. Macleod. On sinusoidal parameter estimation. In Proc. Digital Audio
Eects Workshop (DAFx), Queen Marys, London, 2003.
[27] S.W. Hainsworth and M.D. Macleod. Onset detection in musical audio signals. In Proc. International
Computer Music Conference, Singapore, 2003.
[28] S.W. Hainsworth, M.D. Macleod, and P.J. Wolfe. Analysis of reassigned spectrograms for musical
transcription. In IEEE Workshop on Applications of Signal Processing to Audio and Acoustics,
Mohonk, NY, October 2001.
26
[29] V.K. Jain, W.L. Collins, and D.C. Davis. High-accuracy analogue measurements via interpolated
FFT. IEEE Trans. Instr. and Meas., IM-28(2):11322, June 1979.
[30] F. Keiler and S. Marchand. Survey on extraction of sinusoids in stationary sounds. In Proc. Digital
Audio Eects Workshop (DAFx), pages 518, 2002.
[31] F. Keiler and U. Zolzer. Extracting sinusoids from harmonic signals. J. of New Music Research,
30(3):24358, 2001.
[32] K. Kodera, C. de Villedary, and R. Gendrin. A new method for the numerical analysis of non-
stationary signals. Physics of the Earth and Planetary Interiors, 12:14250, 1976.
[33] K. Kodera, R. Gendrin, and C. de Villedary. Analysis of time-varying signals with small BT values.
IEEE Trans. ASSP, 26(1):6476, February 1978.
[34] M.D. Macleod. Fast nearly ML estimation of the parameters of real or complex single tones or
resolved multiple tones. IEEE Trans. Signal Processing, 46(1):141148, January 1998.
[35] M.D. Macleod. High resolution nearly-ML estimation of sinusoids in noise using a fast frequency
domain approach. In Proc. EUSIPCO, pages 184952, 1998.
[36] S. Marchand. Improving spectral precision with an enhanced phase vocoder using sig-
nal derivatives. In Proc. Digital Audio Eects Workshop (DAFx), 1998. available on-line:
http://www.iua.upf.es/dafx98/papers/.
[37] J.E. Odegard, R.G. Baraniuk, and K.L. Oehler. Instantaneous frequency estimation using the
reassignment method. In Proc. 68th SEG Meeting, New Orleans, USA, 1998.
[38] G. Peeters and X. Rodet. Non-stationary analysis/synthesis using spectrum peak shape distortion,
phase and reassignment. In Proc. ICSPAT99, 1999.
[39] G. Peeters and X. Rodet. SINOLA: a new analysis/synthesis method using spectrum peak shape
distortion, phase and reassigned spectrum. In Proc. International Computer Music Conference,
1999.
[40] W.J. Pielemeier and G.H. Wakeeld. A high-resolution time-frequency representation for musical
instrument signals. J. Acoust. Soc. Am., 99(4):238296, April 1996.
[41] F. Plante and W. A. Ainsworth. Formant tracking using reassigned spectrum. In Proc. Eurospeech,
pages I:7414, 1995.
[42] F. Plante, G. Meyer, and W.A. Ainsworth. Improvment of speech spectrogram accuracy by the
method of reassignment. IEEE Trans. SAP., 6(3):2827, May 1998.
[43] B.G. Quinn. Estimating frequency by interpolation using Fourier coecients. IEEE Trans. Signal
Processing, 42(5):12648, May 1994.
[44] B.G. Quinn. Estimation of frequency, amplitude and phase from the DFT of a time series. IEEE
Trans. Signal Processing, 45(3):8147, March 1997.
[45] C. Richard and R. Lengelle. Joint recursive implementation of time-frequency representations and
their modied version by the reassignment method. Signal Processing, 60:16379, 1997.
[46] D.C. Rife and R.R. Boorstyn. Multiple tone parameter estimation from discrete-time observation.
Bell System Technical Journal, 55(9):13891410, November 1976.
[47] X. Serra. Musical sound modeling with sinusoids plus noise. In C. Roads, S. Pope, A. Picialli, and
G. De Poli, editors, Musical Signal Processing, chapter 3, pages 91122. Swets and Zeitlinger, 1997.
[48] McGill University. Mcgill university master samples collection.
http://www.music.mcgill.ca/resources/mums/html/mums.html.
[49] J. Ville. Theorie et applications de la notion de signal analytique. Cables Transmission, 2A:6174,
1948.
[50] P.J. Walmsley. Signal separation of musical instruments: 1st year PhD report. Cambridge University,
1997.
[51] E.P. Wigner. On the quantum correction for thermodynamic equilibrium. Physics Review, 40:749
59, 1932.
27