Note : Les descriptions sont présentées dans la langue officielle dans laquelle elles ont été soumises.
CA 02209916 1997-07-08
WO 96/21926 PCTIUS96/00350
A HARMONIC AND FREQUENCY-LOCKED LOOP PITCH TRACKER
AND SOUND SEPARATION SYSTEM
The present invention relates generally to pitch tracking systems, methods
for tracking the pitch of a quasi periodic sound source and for the separation
of
periodic signals from mixtures of sounds.
,EACKGROUND OF THE INVENTION
Pitch tracking is of interest whenever a single quasi periodic sound source is
to be studied or modeled. For instance, the trajectory of a sound's pitch,
also
called the fundamental frequency, over a period of time can also be used to
synthesize similar or related sounds using speech or musical synthesis
techniques.
An example of a quasi periodic sound source is a singer's voice singing a
particular
note (e.g., high C). The sound generated by the singer typically has a certain
amount of vibrato or pitch modulation, noise and aperiodicity in the wave
shape,
making the sound quasi periodic rather than a pure periodic signal.
Currently pitch detection methods can be classified into three categories:
Fourier-based frequency domain techniques, time domain techniques, and methods
which use both techniques. The present invention is a time domain technique.
In time domain "feature detection methods", the input signal is usually
preprocessed to accentuate some time domain feature, and the time between
occurrences of that feature is calculated as the period of the signal. The
pitch and
the period of the input signal are related by the equation: pitch = 1/period.
A
typical time domain feature detector includes a low pass filter for detecting
peaks
or zero crossings of the filtered signal. Since the time between occurrences
of a
CA 02209916 1997-07-08
WO 96/21926 I'CT/US96/00350
- 2 -
particular feature is used as the period estimate, feature detection schemes
usually
do not use all of the data available. Selection of a different feature often
yields a
different set of pitch estimates. Since estimates of the period are often
defined at
the instant when the features are detected, the frequency samples yielded are
not
uniformly distributed in time. To avoid the problem of non-uniform time
sampling,
a window of fixed sized can be moved through the signal in order to obtain an
averaged period estimate.
Other prior art time domain methods include the use of auto correlation
functions or difference norms to detect the similarity between the wave form
and a
time lag version of itself. However, prior art methods were computationally
inefficient, with real time performance infeasible.
SUMMARY OF THE INVENTION
In summary, the present invention is a system and method for tracking the
pitch of a quasi periodic signal in a mixture of signals. The quasi periodic
signal
is "frequency warped" by selectively frequency modulating it, thereby
resulting in a
signal that is stationary and is a simplified spectrum which is more amenable
to
analysis. The resultant demodulated signal is low pass filtered resulting in
an
analytic signal whose phase winding rate is the frequency mismatch error
between
the target signal and the demodulating signal. The phase is differenced by
multiplying the signal with a delayed version of itself creating an
instantaneous
autocorrelation. Thereafter the phase difference is measured with a complex
arctangent to yield a resulting phase error. The resulting phase error is
input to an
integrator whose output value is the estimate of the frequency. This output
frequency parameter is then used to update the demodulating signal thus
closing the
signal loop.
In a second embodiment of the present invention, a plurality of frequency
locked loop trackers are servoed together centering each one of the trackers
on a
multiple of the fundamental frequency of the input signal. The resulting phase
errors derived from the frequency lock loop trackers are weighted to improve
CA 02209916 2005-02-23
74673-10
3
system performance. In one embodiment, the frequency
corrections from each tracker are weighted with the inverse
variance of its tracking performance. Accordingly,
harmonics with low variance are weighted strongly, and
harmonics in a noisy region of the spectrum and thus high
variance will be weighted less strongly. The resulting
fundamental frequency estimate is a minimum-variance
estimate, and is better than the best single frequency
locked loop estimate. The weighted phase error is then fed
back to an integrator to yield a high resolution estimate of
the target signal fundamental frequency and all of its
harmonics. The amplitude envelopes for each partial signal
can be easily extracted arid used in conjunction with the
fundamental estimate from each frequency lock loop tracker
to resynthesize the signal in isolation from the mixture.
Since the resynthesized signal is in phase with the original
signal, the target may be removed from the mixture by
subtraction.
According to one aspect the invention provides a
frequency-locked loop pitch tracker for tracking an input
signal comprising: demodulation means including a
demodulation signal for demodulating said input signal
resulting in a complex demodulated signal; a low pass filter
receiving said complex demodulated signal, said low pass
filter for producing a filtered analytic signal; means for
detecting a rate of phase change of said filtered analytical
signal and for producing a frequency tracking error signal;
an accumulator for receiving said frequency tracking error
signal and outputting an estimated input signal frequency;
and means for updating said demodulation signal responsive
to said estimated input signal frequency; said accumulator
including an integrator for receiving said frequenr-y
tracking error signal and producing an integrator output
CA 02209916 2005-02-23
~ =
74673-10
3a
signal, and a frequency-smoothing filter coupled to said
integrator for receiving said integrator output signal and
producing an improved frequency estimate signal.
According to another aspect the invention provides
a frequency-locked loop pitch tracker for tracking an input
signal by tracking a plurality of harmonics in a harmonic
signal representation of said input signal comprising: a) a
like plurality of frequency trackers, each of said frequency
trackers for tracking one of said harmonics, each of said
frequency trackers including demodulation means including a
demodulation signal for demodulating said one of said
harmonics resulting in a complex demodulated signal; a low
pass filter receiving said complex demodulated signal, said
low pass filter for producing a filtered analytic signal;
means for detecting a rate of phase change of said filtered
analytical signal and for producing a frequency tracking
error signal; wherein said plurality of frequency trackers
are harmonically constrained such that each frequency
tracker tracks a respective integer multiple of a
fundamental frequency component of said input signal;
wherein said each of said frequency trackers further
includes a variance estimator for calculating the variance
of said frequency tracking error signal; b) means for
weighting each of said frequency tracking error signals from
each of said plurality of frequency trackers for producing a
weighted frequency tracking error signal; c) an accumulator
for receiving said weighted frequency tracking error signals
and outputting an estimated input signal frequency; and d)
means for updating said demodulation signal responsive to
said estimated input signal frequency.
According to another aspect the invention provides
a pitch tracker for tracking an input signal by tracking a
plurality of harmonics in a harmonic signal representation
CA 02209916 2005-02-23
74673-10
3b
of said input signal comprising: a) a like plurality of
frequency trackers, each of said frequency trackers
responsive to an estimated frequency signal for tracking one
of said harmonics and producing a frequency tracking error
signal; wherein said plurality of frequency trackers are
harmonically constrained such that each frequency tracker
tracks a respective integer multiple of a fundamental
frequency component of said input signal; wherein said each
of said frequency trackers further includes a variance
estimator for calculating the variance of said frequency
tracking error signal; b) means for weighting each of said
frequency tracking error signals from each of said plurality
of frequency trackers for producing a weighted frequency
tracking error signal; wherein each respective one of said
frequency tracking error signals is weighted in accordance
with the inverse of the variance of said respective
frequency tracking error signal; and c) an accumulator for
receiving said weighted frequency tracking error signals and
outputting an updated estimated frequency signal such that
each said frequency tracker tracks a corresponding one of
said harmonics in accordance with said updated frequency
estimate signal.
According to another aspect the invention provides
a frequency-locked loop method for tracking an input signal
comprising the steps of: demodulating said input signal
with a demodulation signal resulting in a complex
demodulated signal; filtering said complex demodulated
signal with a low pass filter, said low pass filter for
producing a filtered analytic signal; detecting a rate of
phase change of said filtered analytical signal to produce a
frequency tracking error signal; outputting an estimated
input signal frequency responsive to said frequency tracking
error signal; and updating said demodulation signal
CA 02209916 2005-02-23
.s .
74673-10
3c
responsive to said estimated input signal frequency; said
outputting step including integrating said frequency
tracking error signal to produce an integrator output
signal, and filtering said integrator output signal with a
frequency-smoothing filter to produce an improved frequency
estimate signal.
According to another aspect the invention provides
a method for tracking an input signal by tracking a
plurality of harmonics in a harmonic signal representation
of said input signal comprising: a) providing a like
plurality of frequency trackers, each of said frequency
trackers demodulating said input signal with a demodulation
signal for tracking one of said harmonics; wherein said
plurality of frequency trackers are harmonically constrained
such that each frequency tracker tracks a respective integer
multiple of a fundamental frequency component of said input
signal; b) deriving a frequency error tracking signal for
each of said harmonics; c) weighting each of said frequency
tracking error signals from each of said plurality of
frequency trackers for producing a weighted frequency
tracking error signal; d) outputting an estimated input
signal frequency responsive to said weighted frequency
tracking error signal; and e) updating said demodulation
signal responsive to said estimated input signal frequency;
further including the step of determining the variance of
said frequency tracking error signal for each of said
harmonics, according to the formula
E2k[n] = gk[n] E2k[n-1]+(1-gk[n] )EZk[n]
where
E2k[n] is the variance estimate;
CA 02209916 2005-02-23
74673-10
3d
ek[n] is the frequency tracking error signal for
kth harmonic,
and gk[n] is the loop gain.
According to another aspect the invention provides
a method for tracking an input signal by tracking a
plurality of harmonics in a harmonic signal representation
of said input signal comprising: a) providing a like
plurality of frequency trackers, each of said frequency
trackers demodulating said input signal with a demodulation
signal for tracking one of said harmonics; wherein said
plurality of frequency trackers are harmonically constrained
such that each frequency tracker tracks a respective integer
multiple of a fundamental frequency component of said input
signal; b) deriving a frequency error tracking signal for
each of said harmonics; c) weighting each of said frequency
tracking error signals from each of said plurality of
frequency trackers for producing a weighted frequency
tracking error signal; d) outputting an estimated input
signal frequency responsive to said weighted frequency
tracking error signal; and e) updating said demodulation
signal responsive to said estimated input signal frequency;
further including the steps of determining a variance
estimate of said frequency tracking error signal for each of
said harmonics, and determining when said variance estimate
saturates; said weighting step including limiting the
weighting of each frequency tracking error signal whose
variance estimate saturates.
According to another aspect the invention provides
a method for tracking an input signal by tracking a
plurality of harmonics in a harmonic signal representation
of said input signal comprising: a) providing a like
plurality of frequency trackers, each of said frequency
CA 02209916 2005-02-23
74673-10
3e
trackers demodulating said input signal with a demodulation
signal for tracking one of said harmonics; wherein said
plurality of frequency trackers are harmonically constrained
such that each frequency trackers tracks a respective
integer multiple of a fundamental frequency component of
said input signal; b) deriving a frequency error tracking
signal for each of said harmonics; c) weighting each of said
frequency tracking error signals from each of said plurality
of frequency trackers for producing a weighted frequency
tracking error signal; d) outputting an estimated input
signal frequency responsive to said weighted frequency
tracking error signal; and e) updating said demodulation
signal responsive to said estimated input signal frequency;
further including the step of determining the variance of
said frequency tracking error signal for each of said
harmonics; wherein said weighting step includes: a)
weighting each of said frequency tracking error signals by
the reciprocal of said variance determined for each of said
frequency tracking error signals; and b) summing all of the
weighted frequency trackirig error signals to yield said
weighted frequency tracking error signal.
BRIEF DESCRIPTION OF THE DRAWINGS
Additional objects of interest to the invention
will be more readily apparent from the following description
and appended claims when taken in conjunction with the
drawings, in which:
Figure 1 is a frequency locked loop tracker
according to the preferred embodiment of the present
invention.
Figure 2 shows the frequency locked loop tracker
of Figure 1 including a phase locked loop.
CA 02209916 2005-02-23
74673-10
3f
Figure 3 shows the frequency locked loop tracker
of Figure 1 including an improved frequency estimation means
outside the tracking loop.
Figure 4 is a frequency locked loop tracker
according to the preferred embodiment of the present
invention including a resynthesis module.
Figure 5A shows the frequency locked loop tracker
of Figure 4 including a delay line for compensating for the
low pass filter group delay.
Figure 5B shows the frequency locked loop tracker
of Figure 5A including a subtraction module for removing the
resynthesized partial signal from the input signal.
Figure 6A is a frequency locked loop tracker
according to Figure 3
CA 02209916 1997-07-08
WO 96/21926 PCT/US96/00350
- 4 -
including a resynthesis module.
Figure 6B shows the frequency locked loop tracker of Figure 6A including
a subtraction module for removing the resynthesized partial signal from the
input
signal.
Figure 7 is a harmonic locked loop tracker in which a plurality of frequency
locked loop trackers according to the preferred embodiment of the present
invention are servoed for tracking a partial signal and a plurality harmonics
of the
partial signal.
DESCRIPTION OF THE PREFERRED EMBODIMENTS
Referring to Figure 1, the pitch track of the present invention 100 is shown.
The pitch tracker 100 receives as an input signal z[n] 102 which is a mixture
of a
p[n] complex valued discrete time signal and some unknown disturbance signal
v[n] wherein
z [n] = p [n] + v [n]
The target signal p[n] is a complex value discrete time signal defined for n>0
with
a sampling frequency f S wherein
p[n] = a[n] exp j2,, )
E f[k] + J4)o
k=1
where a[n] is the instantaneous amplitude envelope,
f[n] is the instantaneous frequency, and
0o is the phase offset at time n=0.
The first step in the analysis of the input signal z[n] 102 is to demodulate
the input
signal by means of a frequency matched demodulation signal. In particular, the
=
input signal z[n] 102 is demodulated by multiplier 104, which multiplies the
input
signal z[n] with the complex conjugate of a frequency warping signal -"-:'[n]
106.
CA 02209916 1997-07-08
WO 96/21926 PCT/US96/00350
- 5 -
The use of the frequency warping signal 106 allows for the elimination of the
FM
band width component due to the instantaneous frequency modulation of the
carrier. The frequency warping signal 106 demodulates the input signal z[n]
102
by means of a signal which is frequency matched to the input signal z[n] 102.
In
the preferred embodiment of the present invention, the input signal z[n] is
demodulated using a complex phasor which rotates at a frequency equal to a
frequency estimate generated by the pitch tracker 100. The frequency matching
will be described in greater detail below in conjunction with the frequency
estimate
generated by the pitch tracker of the present invention. For the purposes of
this
first step of the analysis, it will be assumed that a frequency matched
demodulation
signal is provided. Those ordinarily skilled in the art will recognize that if
the
frequency estimate is equal to the target frequency, then the frequency
matched
demodulation by the instantaneous frequency f(t) of the estimate signal will
yield a
constant phase signal d[n] at or near DC.
The second step of the analysis requires low pass filtering of the constant
phase signal to improve the signal to noise ratio. In particular, the complex
demodulated signal d[n] resulting from the multiplication of the input signal
z[n]
102 with the complex conjugate of the frequency warping signal 106 is coupled
to
a low pass filter 108. The low pass filter 108 improves the signal to noise
ratio by
low pass filtering the demodulated signal d[n] thereby attenuating the
demodulated
noise portion of the input signal.
In the preferred embodiment of the present invention, the low pass filter has
a cut off frequency of fc and unity gain at DC. The low pass filter may be of
time-varying or time-invariant form with a fixed fc. A time-varying filter can
be
used with a dynamically adjustable bandwidth wherein a wide cut-off frequency
is
programmed before frequency lock is achieved, and thereafter bandwidth can be
reduced. However, dynamically altering the filter characteristics may
introduce
artifacts into the filter output if changes are made suddenly. Accordingly, in
the
preferred embodiment of the present invention, a time-invariant filter with a
wide
bandwidth is utilized providing a wide frequency lock-in range. A typical cut-
off
CA 02209916 1997-07-08
WO 96/21926 PCT/US96/00350
- 6 -
frequency would be 50-100 Hz. Wider cut-off frequencies are beneficial for
tracking signals with rapidly varying frequency modulation, whereas narrower
cut-
off frequencies allow for better noise rejection.
In the next step of the analysis, the resultant low pass filtered signal is
sampled to measure the phase difference of the filtered signal. The resultant
signal
u[n] is multiplied by means of multiplier 110 with a delayed and complex-
conjugated version of itself via delay line 112. The change in phase of the
resultant signal u[n] from the low pass filter 108 is then calculated by using
a
standard argument function 114 in order to result in the change in phase
~~Jn].
The frequency tracking error at time [n] is thereafter defined as E f[n] where
E f [n] ~o~u [n]
Accordingly the change in phase 0~õ[n] is normalized by multiplying the change
in
phase signal by the sampling frequency divided by 2n (fs/27c) by multiplier
116
and results in an instantaneous frequency tracking error at time [n]. Note
that the
scaling factor may be left off resulting in calculations in radians per sample
as
opposed to hertz. In the preferred embodiment of the present invention the
sampling frequency is 44,100 Hz, however, other sampling frequencies as is
known
in the art may be utilized. The frequency tracking error represents the error
between the frequency estimate (generated by the pitch tracker 100 for use in
demodulating the input signal z[n]) and the frequency of the target signal
p[n].
Having calculated the frequency tracking error, the pitch tracker 100 utilizes
this error information to generate a better frequency estimate for use in
demodulating the input signal. Specifically, the frequency tracking error E
f[n] is
combined with an attenuation tracking gain signal g[n] by multiplier 118 for
input
into integrator 120. The gain signal g[n] controls how fast the system will
adapt to
the particular frequency error E f[n]. The combination of the frequency error
E f[n]
and the gain signal g[n] yields an attenuated frequency error signal. The
attenuated
CA 02209916 1997-07-08
WO 96/21926 PCTIUS96/00350
- 7 -
frequency error signal is coupled to an integrator 120 in order to derive the
estimated frequency output Y[n] for use in updating the demodulation signal.
Those ordinarily skilled in the art will recognize that any filtering or
smoothing
means may be used as is known in the art in lieu of the simple attenuated
frequency integrator. In the preferred embodiment the integrator output, which
reflects the estimated frequency of the target signal, must be initialized for
tracking
a particular desired partial signal. This may be accomplished by providing a
particularized user input associated with the frequency of a particular
partial signal
to be tracked or may be accomplished by performing a sweep over an audio band
in order to isolate a particular partial signal. Alternatively, a peak-
detection
scheme may be used on a FFT of an initial segment of the input signal to find
a
candidate initial frequency. Those ordinarily skilled in the art will
recognize that
the frequency tracker 100 will naturally track the strongest sinusoidal in the
pass
band of the low pass filter, and accordingly, the accuracy of the initial
frequency
estimate is not critical.
Finally the loop is closed by providing the frequency estimate to a phase
accumulator for updating the frequency warping signal for use in demodulating
the
input signal. Specifically, the integrator estimated frequency output Y[n]
from
integrator 120 is scaled via multiplier 122 by combining the estimated
frequency
with a scaling signal (21t/fs where fs is the sampling frequency). The scaled
output
is coupled to a phase accumulator 124 for use in deriving an estimated phase
responsive to the estimated frequency Y[n]. The estimated phase is then used
as
the estimated phase of the demodulating phasor to produce the warping signal
106
for use in the demodulation of the input signal z[n]. The phase accumulator
124
includes an integrator which derives an estimated phase from the scaled
estimated
frequency provided from the integrator 120. The derived phase is the estimated
phase of the demodulating phasor for use in demodulating the input signal
z[n]. In
the preferred embodiment, this is accomplished by transforming the estimated
phase
into a sinusoid by taking the cosine and sine of the phase to generate a
complex
sinusoidal signal. Additionally, the phase is wrapped in a periodic fashion in
order
to prevent overflow of the phase accumulator 124.
CA 02209916 1997-07-08
WO 96/21926 PCT/US96/00350
- 8 -
Those ordinarily skilled in the art will recognize that the combination of the
output estimate frequency from the integrator 120 in conjunction with the
scaling
multiplier 122 and the modulator 124 for deriving a frequency warping signal
106
is equivalent to a voltage controlled oscillator wherein the input frequency
is used =
to derive a frequency matched demodulation signal. As such, the description of
the
integrator and phase accumulator according to the preferred embodiment should
not
be construed as limiting.
Referring now to Figure 2, the frequency locked loop tracker of the present
invention is shown including a phase-locked loop for more feedback control. In
this embodiment, a phase-locked loop is provided for locking to the phase of
the
demodulated and filtered signal u[n] described in conjunction with the first
embodiment above. In the preferred embodiment described above, the frequency
of
a target signal is tracked but the phase is not. By providing a phase-lock
feedback
term, phase lock as well as frequency lock may be attained. The extra phase
information provides for better isolation of the target signal for subtractive
analysis.
In this embodiment, the pitch tracker is more sensitive to noise and phase
locking
is difficult to attain in rapidly changing signals. Again, the analysis begins
by
demodulating a complex input signal z[n] 102 via multiplier 104 by a frequency
warping signal 106 resulting in the complex demodulated signal d[n]. The
complex demodulated signal d[n] is coupled to a low pass filter 108 producing
an
analytic output u[n].
The analytic signal u[n] is used in achieving phase lock by adding a
modification to the frequency lock method described in the preferred
embodiment.
The phase lock loop is created by providing a second loop for tracking the
phase
mismatch error between the frequency warping signal 106 and the input signal
z[n]
102. This is accomplished by taking the argument 202 of the analytic signal
u[n]
which yields a phase error. The resultant phase error is attenuated by a phase
gain
signal g,[n] via multiplier 204. The resultant attenuated phase error signal
is
coupled to the phase accumulator 124 of the preferred embodiment. Internal to
the
phase accumulator 124, this attenuated phase error is combined via an internal
CA 02209916 1997-07-08
WO 96/21926 PCT/US96/00350
_ g _
integrator with the derived phase estimate for phase lock. Those ordinarily
skilled
in the art will recognize that there are now two competing forces trying to
guide
the tracking. Close attention must be paid to the relative ratios of the gain
gn and
the phase gain g,[n] since both phases range over [-Tt, 7t]. Accordingly, g[n]
must
be much greater than go[n]. However, as frequency lock is obtained, the phase
gain g,[n] can be varied to be large enough to ensure that quick phase
tracking
convergence occur. Those ordinarily skilled in the art will recognize that
automatic
gain control algorithms which track the status of the frequency lock can
adjust the
gain g[n] and phase gain g,[n] making them dependant on the variances in the
phase difference 0ou[n] and the phase mismatch error ~u.
Referring now to Figure 3, the present invention is shown including a
second frequency estimate }'t[n-8I-82] for providing a frequency estimate
including group delay compensation outside the "loop" for use in resynthesis
or
other means as is known in the art. The basic tracking loop is identical to
that
shown in Figure 1, however, a second frequency estimate is made outside of the
loop based on the crude estimates of }'[n] from a first pass of a partial
signal to be
tracked along with the error estimation updates f[n]. The crude estimates
are then,
refined using a Kay optimal phase-difference smoother.
Specifically, the estimated frequency Y[n] output from the integrator 120 is
coupled via a delay line 304 to the frequency error signal ef[n] via adder
306.
Since the new estimate is made outside the loop, the new estimate does not
contribute to tracking dynamics. The group delay of the low pass filter 108 is
taken into account by the delay line 304. The output of the adder 306, which
is
effectively the phase difference of the input signal if it had not been
demodulated
by the frequency warping signal 106, is then coupled to a Kay smoother 302
having a group delay of 52. In the preferred embodiment, the Kay smoother 302
is simply an FIR filter with quadratic coefficients given by the formula
Wkay [ n] = 6 N Jn_ (n)2}
N2-1 N N
CA 02209916 1997-07-08
WO 96/21926 PCT/US96/00350
- 10 -
for 1 < n :!~N-1.
The Kay smoother output then reflects an improved estimate of the frequency
being
tracked. This improved estimate Yt[n-81-52] may be used in providing a
resynthesized partial signal as will be described below.
Referring now to Figure 4, the frequency locked loop tracker 100 of the
preferred embodiment of the present invention is shown including a resynthesis
module 401. Often it may be desired to produce a resynthesized partial signal
p[n]
which is a cleaned up version of the partial signal p[n] being tracked from
the
input signal z[n]. The cleaned up signal may be derived by combining the
frequency warping signal 106 with the analytic signal u[n] via multiplier 402.
The
resultant output of this combination is an estimated partial signal p[n] which
reflects the combination of the estimated frequency from the integrator 120
(as
embodied in the frequency warping signal 106) combined with the envelope
signal
u[n].
Those ordinarily skilled in the art will recognize that this frequency locked
loop tracker does not compensate for the group delay of the low pass filter
108. A
better estimation of the partial signal p[n-81] can be derived by providing a
delay
line 502 as shown in Figure 5A. The delay line 502 provides compensation for
the '
group delay of the low pass filter and accordingly provides a more accurate
resynthesized partial signal. Specifically, the delay line 502 couples the
frequency
warping signal 106 to the multiplier 402 yielding an improved estimate that
accounts for the group delay of the low pass filter.
In addition to the isolation of a particular partial'signal from a given input
signal as described above, it is often desirous to produce a filtered input
signal
which has had the target signal removed. Examples of applications where this
may
be used is in the removal of a "voice" or musical instrument from a musical
selection (e.g. audio signal) or the removal of background noise from a
"voice".
This process is known as notch-filtering, and when applied will result in a
notch-
CA 02209916 1997-07-08
WO 96/21926 PCT/US96/00350
- 11 -
filtered output signal. In the preferred embodiment, the partial signal p[n]
or
p[n-81] may be used in a notch-filter process to derive a notch-filtered
output
signal as shown in Figure 5F. The notch-filtered output signal is derived by
subtracting the resynthesized partial signal p[n] from the input signal z[n].
In the
preferred embodiment, the input signal z[n] is coupled via a second delay line
504
.to a first input of a subtractor 506. The second input of the subtractor 506
receives
the resynthesized partial signal p[n-Sl] from above. The subtractor 506
outputs a
notch-filtered signal resulting from the subtraction of the partial signal
from the
input signal.
Referring now to Figure 6A, a second resynthesis module 601 for
resynthesizing a partial signal is shown. The basic frequency locked loop
tracker
of Figure 1 is included with the Kay smoother filter of Figure 3 in order to
make
use of the improved frequency estimate Yt[n-8l-82] in producing a
resynthesized
partial signal. Specifically, the improved frequency estimate ft[n-S1-S2] is
scaled
by combining it with a scaling signal (2trl fs where f S is the sampling
frequency)
via multiplier 604. The scaled frequency is then coupled to a second phase
accumulator 602 which integrates the scaled frequency to create an improved
estimated phase of the demodulating phasor for the phase accumulator 602. The
phase accumulator 602 outputs a second frequency warping signal 606 which is
utilized in demodulating a delayed version of the input signal zn. This is
accomplished by coupling the input signal zo via delay line 608 to multiplier
610
for combining with the second frequency warping signal 606.
The complex demodulated signal dt[n-S1-S2] is then coupled to a second
low pass filter 612 having a group delay of 83. The output of the second low
pass
filter 612 is coupled with the second frequency warping signal 606 via
multiplier
614 in order to yield an improved partial signal pt[n-81-82-83]. The second
low
pass filter is the resynthesis filter, and is designed to allow for higher-
quality
filtering characterized by a narrower cut-off frequency and linear phase
response.
Those ordinarily skilled in the art will recognize that a delay line 616 may
be used
to couple the second frequency warping signal 606 to the multiplier 614 in
order to
CA 02209916 1997-07-08
WO 96/21926 PCT/US96/00350
- 12 -
account for the group delay of the second low pass filter 612. Accordingly,
the
resultant output of the combination of the delayed second frequency warping
signal
606 and the analytic signal from the low pass filter 612 will result in an
improved
partial signal pt[n-51-82-53]. Because this resynthesized signal is generated
outside
the normal tracking loop, no tracking dynamics will be affected by this
resynthesis
function. Those ordinarily skilled in the art will recognize that the more
efficient
estimate of the partial signal p[n] can be used to calculate a high quality
notched
filter signal as is known in the art.
Again, the partial signal p[n-Si-S2-S3] may be used in a notch-filter process
to derive a notch-filtered output signal as shown in Figure 6B. The notch-
filtered
output signal is derived by subtracting the resynthesized partial signal p[n]
from the
input signal z[n]. In the preferred embodiment, the input signal z[n] is
coupled via
a fourth delay line 618 to a first input of a subtractor 620. The second input
of the
subtractor 620 receives the resynthesized partial signal p[n-51-52-53] from
above.
The subtractor 620 outputs a notch-filtered signal resulting from the
subtraction of
the partial signal from the input signal.
Referring now to Figure 7, a plurality of frequency locked loop trackers
700-1 to 700-N according to the preferred embodiment of the present invention
are
servoed in a harmonic locked loop tracker 701. The frequency locked loop
tracker
of the preferred embodiment of the present invention performs fast and
accurate
tracking of the instantaneous frequency of a single target partial signal in
isolation.
However if the signal to noise ratio is large, tracking may break down.
Acoustical
signals are often composed of complex mixtures of signals which bring the
signal
to noise ratio for a target partial signal down below the level rieeded for
tracking
according to the frequency locked loop method disclosed above. However, the
harmonic structure of many natural acoustic signals allows for the robust
tracking
of the harmonic set of partials associated with a given harmonic signal.
Accordingly, a harmonic locked loop tracker 701 is provided wherein a
plurality of
frequency locked loop trackers are servoed to track a partial signal and a
plurality
of harmonics where each of the harmonics is a multiple of the fundamental
CA 02209916 1997-07-08
WO 96/21926 PCT/US96/00350
- 13 -
frequency of the partial signal being tracked.
In the first step of the analysis of a harmonic signal s[n], an instantaneous
frequency correction term is calculated for each harmonic. Specifically, the
harmonic signal s[n] is demodulated by the frequency warping signal 706 via
multipliers 704 for each stage. Each stage further includes a low pass filter
708
which receives the complex demodulated signal dk[n] which in turn produces an
analytic signal uk[n]. This resultant signal uk[n] is then combined with a
conjugate
of itself delayed by one sample via multiplier 710 and delay element 712. The
resultant output of the multiplier 710 is coupled to a phase extraction module
714
in order to calculate the phase difference of the resultant signal. The phase
extraction module 714 is normalized by combining a normalization signal
(fs/2Ttk
where fs is the sampling frequency) via multiplier 716, resulting in a error
term
tf~`o[n]. The division by "k" takes into account that the kth stage is
tracking "k"
times the fundamental frequency.
In the second step of the analysis, the resulting error signals ~ko[n] are
combined for each stage to yield an overall optimized error correction for use
by
the frequency estimator and phase accumulator of the frequency locked loop
tracker
disclosed above. In the preferred embodiment, the frequency corrections from
each
tracker are weighted in accordance with the inverse of the variance of its
tracking
performance. Hence each harmonic of the tracked fundamental signal with a low
variance will be weighted strongly, while harmonics with high variance (e.g.,
in
noisy portions of the spectrum) will be weighted less strongly. The resultant
fundamental frequency estimate is a minimum variance estimate, and is better
than
the best single frequency locked loop estimate.
Specifically, the error signal i~ko[n] is utilized in order to calculate a
variance estimate for each of the individual phase trackers. In each tracker,
the
error signal ei(ko[n] is multiplied by itself via squaring module 750. The
output of
the squaring module 750 is coupled to a variance estimator 752 utilized to
calculate
the variance of the error signal E~(ko[n]. The variance estimator 752 derives
a
CA 02209916 1997-07-08
WO 96/21926 PCTIUS96/00350
- 14 -
variance estimate z fko[n] according to the formula
Z fko[n] = gk[ri]ZUkp[n-1] + (l-gk[n])(~',`o[n])2
wherein the time constant gk[n] may be time varying and an exponential
weighting
scheme is used. Those ordinarily skilled in the art will recognize that other
weighting schemes may be utilized in order to determine how the individual
phasor
signals will be combined in order to optimize partial signal tracking.
In the preferred embodiment of the present invention, the resultant variance
estimate Z fko[n] is inverted by module 754 and then coupled to a saturation
detector 756. The saturation detector serves to compensate for signals with a
high
signal to noise ratio for the particular harmonic being tracked. When the
signal to
noise ratio is too high, the variance estimate becomes limited by the band
width of
the low pass filter 708 causing it to be too low. When the variance estimate
is
saturated in this way, it causes the weighting for its associated tracker to
be too
high. This saturated variance estimate associated with the particular harmonic
tracking stage then becomes an unreliable estimator of the true variance of
the
single target partial p[n] for this particular harmonic. This is especially a
problem
for higher harmonics where often a mix of broad band noise and audio signals
occurs. The weighting given to the particular frequency and phase error
associated
with the individual harmonic is proportional to the reciprocal of the
estimated
variance thus not allowing for the higher harmonics to become unfairly highly
weighted. In the preferred embodiment, the saturation detector 756 output
wk[n] is
'25 defined as
wk[n] = 1/tuko[n] if tuko[n]< BW2/24k2
otherwise wk[n] = 1/k2a fko[n]
where BW equals the bandwidth of the kth low pass filter 708.
The output of the saturation detector is combined via multiplier 757 with
the individual error signal tfko[n] to yield a weighted phase error signal.
Each of
the weighted error signals are combined by adders 758 and combined with the
sum=
CA 02209916 1997-07-08
WO 96/21926 PCT/US96/00350
- 15 -
of the weights from each of the saturation detectors 756 for each harmonic
phase
tracker. The sum of the weights is inverted prior to combination with the sum
of
the phase error signals by inverter 760 in order to provide a normalizing
factor for
the summed phase error signal. The output of the multiplier 762 is the
weighted
phase error signal which is then combined with the tracker attenuation gain
go[n]
and integrated to produce the estimated fundamental frequency Yo[n] for use in
the
demodulation of the input signal 702 as was described in accordance with the
frequency locked loop tracker above.
Those ordinarily skilled in the art will recognize that any of the number of
weighting schemes may be utilized in order to combine the individual phase
error
signals which result from each harmonic loop tracker. The particular inverse
variance method selected should not be construed as limiting.
The input signal s[n] may include several voices, each comprising a
fundamental partial signal and a set corresponding harmonics. The harmonics
tracked by the set of parallel trackers in Figure 7 can be resynthesized so as
to
regenerate one complete "voice". In one preferred embodiment, such resynthesis
is
accomplished using one instance of the resynthesis module (i.e., multiplier
402)
shown in Figure 4 for each of the trackers. Improved resynthesis is
accomplished
in a second preferred embodiment by providing one instance of the resynthesis
module shown in Figure 5 or Figure 6 for each of the trackers in Figure 7.
Those ordinarily skilled in the art will recognize that the harmonic loop
tracker described in the preferred embodiment may also be used for tracking a
well
defined partial signal along with non-integer multiples of the fundamental
frequency. This type of tracking known as inharmonic tracking is especially
useful
in tracking audio signals such as a piano, wherein sounds emanating from a
piano
are composed of stretched partials which are not integer multiples of a
particular
fundamental frequency. Inharmonic tracking is accomplished by defining a
constant inharmonic ratio between the kth partial and the fundamental
frequency.
Such inharmonic frequency ratios may be supplied by a template or may be
CA 02209916 1997-07-08
WO 96/21926 PCTIUS96/00350
- 16 -
adaptively trained. In the preferred embodiment, the tracking of the
inharmonic
partials is the same with the exception that the kth demodulated signal must
be
computed explicitly, instead of in an iterative cascade, since the partials
are no
longer integer multiples of the fundamental frequency.
ALTERNATE EMBODIMENTS
Although the present invention has been described with reference to a few
specific embodiments, the foregoing descriptions are illustrative of the
invention
and should not to be construed as limiting. Various modifications may occur to
those skilled in the art without departing from the true spirit of the scope
of the
invention as defined by the appended claims.
For instance, the minimum-variance weighting method of the present
invention could be used with a set of harmonically constrained peak detectors
in an
FFT-based pitch tracker.