вход по аккаунту



код для вставкиСкачать
Patent Translate
Powered by EPO and Google
This translation is machine-generated. It cannot be guaranteed that it is intelligible, accurate,
complete, reliable or fit for specific purposes. Critical decisions, such as commercially relevant or
financial decisions, should not be based on machine-translation output.
The present invention relates mainly to the design of digital speech precomps, in particular
digital precomp filters, and proposes a speech precomp filter design scheme using a new class of
design criteria. The filter parameters are determined based on the weightings, approximating the
pre-compact filter to fixed non-zero filter components, and approximating a pre-compensated
model response to a reference system response. Preferably for design, the precomp filter
comprises a fixed non-zero filter component and an adjustable compensator component. The
tunable compensator component is determined by optimizing a criterion function that includes
weightings. Frequency- and / or channel-dependent weighting provides a powerful design tool
that provides complete control over the degree and type of compensation performed in different
frequency bands and / or in different sub-channels. [Selected figure] Figure 5
Digital speech precomp
FIELD OF THE INVENTION The present invention relates mainly to digital speech precomp and in
particular to the design of digital precomp filters, producing and compensating one or several
input signals to a speech production system For those that change the dynamic response of the
system. BACKGROUND OF THE INVENTION Audio production or reproduction systems, including
amplifiers, cables and loudspeakers, always have sometimes undesirable effects on the spectral
characteristics of the sound. The reverberation in the room where the device is installed further
changes the sound. Sound reproduction with very high quality can be achieved by using
matching of cables, amplifiers and top quality loudspeakers, but it is cumbersome and very
expensive. The increased computing power of personal computers and digital signal processors
has created new possibilities for changing the sound characteristics in sound generation or
reproduction systems. It is known that the dynamic characteristics of a speech production system
can be modeled by measuring and recording its response to known test signals. The precomp
filter (R in FIG. 1) is placed between the original sound source and the audio equipment. This
filter is designed and implemented to compensate for the measured characteristics of the speech
production system indicated by H in FIG. In particular, it is desirable to approximate the phase
and amplitude response of the compensated system, shown by D in FIG. 1, to a preset ideal
response. In other words, it is necessary to match the compensated speech reproduction y (t) to
the ideal value yref (t) with a given accuracy. The predistortion generated by the precompensator
R cancels the distortion caused by the system H. The resulting audio reproduction has a sound
feature D. Up to the physical limits of the system, at least in theory, it is possible to achieve
excellent sound quality without extreme high-end voice equipment at high cost. The aim of the
design is, for example, to cancel the acoustic echoes caused by the poorly built loudspeaker
cabinet. Another application is to minimize low frequency echoes due to sound effects at various
different locations within the listening room. Digital precomp filters can be applied to multichannel sound generation systems as well as single loudspeakers. As well as producing better
sound, it can be an important element of the design to produce a particular effect.
For example, the generation of virtual sound sources (sound renderings) is associated with the
sound effects of computer games. A device called a graphic equalizer has been around for quite
some time. The device aims to compensate for the frequency response of a speech production
system by changing its gain in a set of fixed frequency bands. There is a mechanism for
automatically adjusting such filters (see, for example, Patent Document 1). Also, as another prior
art, there is a technique of dividing an audio frequency range into different frequency bands and
configuring different compensators within each of these bands (see, for example, Patent
Documents 2 and 3). Such sub-band solutions suffer from inadequate phase compensation,
especially at the band edges. A method has been proposed to handle the important audio
frequency range as one band. This method requires the use and adjustment of filters with very
high adjustable factors. The proposed method is generally based on the adjustment of a FIR
(finite impulse response) filter to minimize the least squares criterion, and the compensated
signal y (t) and the required response yref (t) The deviation between them is measured (see, for
example, non-patent documents 4 to 7, patent documents 8 to 10). This method seems to be
attractive, as there is an easy-to-handle matching algorithm as well as an off-line design
algorithm that can adjust an IIR filter based on a least squares criterion. There have also been
proposals for nonlinear compensators (see, for example, Patent Document 11 and Non-Patent
Document 14). A solution that proposes measuring the room acoustics response and the
loudspeaker response separately has been used in the design of the precompact inverse filter for
the sound reproduction systems of the documents US Pat. This design partially equalizes both
responses. In U.S. Pat. No. 5,956,015, a method is disclosed for applying both FIR and IIR (infinite
impulse response) filters to the compensation of speech systems. This approach is used to reduce
the number of FIR filter parameters required in the compensation filter. [Patent Document 1] US
Patent No. 4,739,513 [Patent Document 2] US Patent No. 5,384,856 [Patent Document 3] US
Patent No. 5,627,899 [Non-patent document 4] P. M. カールソン(Clarkson)、J.
Moorjopoulos and J.A. K.
Hammond, 1985 "Spectral phase and transient equalization for audio systems", J. Am. Speech
Engineering Society (J. Audio Engineering Society), Vol. 33, pages 127 to 131 [Non-Patent
Document 5] P.I. A. ネルソン(Nelson)、H. Hamada, S.I. J. "Adaptive inverse filtering
for stereo sound reproduction" in 1992 by Elliot, IEEE Proceedings on Signal Processing, Volume
40, pages 1621 to 1632 [Non- Patent Document 6] P.I. A. ネルソン(Nelson)、F. In
1996, "Multichannel signal processing techniques in the reproduction of sound" by OrduaBustamante, J.A. Speech Engineering Association (J. Audio Engineering Society), Vol. 44, pp. 973
to 989 [Non-patent document 7] P.I. A. ネルソン(Nelson)、F. オルドゥア−ブスタマ
ンテ(Ordua−Bustamante)、H. Hamada (1995) "Inverse filter design and
equalization zones in multi-channel sound reproduction systems" (IEEE Proceedings on Speech
and Speech Processing, 3rd issue, "Inverse filter design and equalization zones in multichannel
sound reproduction systems") Volume, page 185 to page 192 [Patent Document 8] US Patent
4,683, 590 [Patent Document 9] US Patent 5,727, 066 [Patent Document 10] International
Patent Publication WO 94/24835 [ Patent Document 11 US Pat. No. 5,438,625 [Patent
Document 12] US Pat. No. 5,511,129 [Patent Document 13] Japanese Patent Application No. 080799880 [Patent Document 11] No. 5,600,718 However, all of the existing methods have
important problems in putting them into practical use.
Compensating filters of design schemes available in the prior art generally have sophisticated
operations and severe practical limits. The resulting automatically occurring compensation filters
are often even dangerous to the audio equipment. Because of the danger of producing a power
compensation signal that is too high. There is a need for design techniques and convenient tools
to avoid the aforementioned drawbacks. The present invention overcomes the above problems in
the prior art. It is a main object of the present invention to provide an improved design scheme
for speech precompilation filters. It is another object of the present invention to provide a flexible
and very accurate method for designing such a filter and to obtain a better control range and
amount of compensation by a precomp filter . In this regard, it is desirable, inter alia, to provide a
filter adjustment technique that provides complete control over the amount of compensation
performed in different frequency bands and / or different audio paths. It is also an object of the
present invention to provide a speech precompensator design method and system that can be
easily handled by existing technology and that achieves excellent compensation capabilities with
a limited number of filter parameters. It is. Yet another object of the present invention is to
provide a flexible and efficient method, system and computer program for designing a digital
speech precomp filter. These and further objects of the present invention are defined by the
appended claims. SUMMARY OF THE INVENTION The present invention provides speech
precomp filters that use novel design criteria. The present invention provides that dynamic
system mathematical model and optimization based on digital precomp filter model provide a
powerful tool for filter design, and various kinds of voice equipment by changing the input signal
Based on the recognition that it will improve In essence, the filter parameters approximate the
precomp filter to fixed non-zero filter components while on the other hand, weighting between
approximating the precompensate model response to the reference system response It is
determined based on For design purposes, the precomp filter preferably comprises a fixed nonzero filter component and a tunable compensator component. The fixed filter component is
usually configured by the filter designer or set as a default configuration, and the adjustable
compensator component is determined by optimizing the criterion function including the above
Similar to fixed filter components, the weighting is usually configured by the filter designer or set
as a default configuration. Once the fixed filter component is configured and the adjustable
compensator component is determined, the filter parameters of the precomp filter can be
computed and implemented. In many practical cases, it is beneficial to include at least one
selectable delay element and a bypass component in the fixed filter component. [0015] By
making the weighting frequency-dependent and / or channel-dependent, it is powerful to provide
complete control over the degree and type of compensation performed in different frequency
bands and / or in different sub-channels. Design tools are obtained. Preferably, the criterion
function comprises penalty periods for frequency-dependent and / or weighted channels. This
period penalizes the compensation portion of the precompensator. This kind of frequencydependent and / or channel-dependent weighting simplifies avoiding dangerous
overcompensation and achieves good compensation in the frequency bands and channels
reached without problems. The optimization of the weighted criterion function can be performed
on-line, for example by using recursive optimization or adaptive filtering or as an off-line design
based on a model. This is similar to traditional on-line optimization. Based on optimization to
adjust a stable and causal realizable infinite impulse response (IIR) compensation filter while
using a limited number of filter parameters to provide good compensation capability A
methodology is proposed. While including a limited number of filter parameters, these digital
filters can cause long impulse responses. A compensation filter so designed may have several
inputs and may output an audio channel. And it can be used to compensate not only multiple
channels but also one channel audio equipment. The proposed design principles and structures
are particularly useful for linear dynamic design models and linear precomp filters, but can be
generalized to non-linear design models and non-linear precomp filters. According to another
aspect of the present invention, there is provided a method of designing a precomp filter for
speech, a system and computer program, the precomp filter designed thereby, and the digital
speech signal generated by the precomp filter as well as the signal. An audio system comprising
such a precomp filter.
The invention offers the following advantages:-Tight control of the range and amount of
compensation performed by the precomp filter, providing complete control over the resulting
acoustic response-Problem Avoiding dangerous overcompensation while achieving good
compensation that can be achieved without-achieving good compensation capability while using
a limited number of filter parameters-and-Optimal to produce superior sound quality and
experience results Get an audio system that is pre-compensated. Other advantages and features
provided by the present invention will be appreciated by the following description of
embodiments of the present invention. Embodiments of the present invention will be described
with reference to the drawings. Further objects of the present invention and their advantages will
be further understood by the following description. Sections 1 to 3 describe the linear case, and
Section 4 generalizes the structure and design principles to the problem with nonlinear and
preferably variable system models as well as nonlinear and preferably variable compensators,
and Section 5 finally describes some of the most primitive aspects. Section 1: Linear Model and
Filter Design For a better understanding of the present invention, it may be useful to begin with
describing a general approach to speech precomp filter design. . The speech production or
reproduction system to be modified is usually represented by a linear invariant dynamic model H.
This model describes the discrete time relationship between a set of p output signals u (t) to a set
of m output signals y (t). Where t represents a discrete time index, and ym (t) (index m is a
“measurement value”. Is an m-dimensional column vector representing the time series of
sound at m different locations, e (t) is the noise, reflection in the unmodeled room, the effect of
incorrect model structure, Show non-linear distortion and other unmodeled contributions. The
operator H is an m × p-matrix, the elements of which are stable linear dynamic operators or
transformations implemented as eg FIR filters or IIR filters.
These filters determine the response y (t) for any input time series vector u (t) in p dimension. A
linear filter or model is represented by such a matrix. This matrix is in the following called
transfer function matrix or dynamic matrix. The transfer function matrix H represents the effects
of the whole or part of the sound generation or reproduction system, and any existing digital
compensators, digital to analog converters, analog amplifiers, loudspeakers, cables and some
applications Also includes room acoustic response. In other words, the transfer function matrix H
represents the dynamic response of the relevant part of the speech production system. The input
signal u (t) to this system is a p-dimensional column vector, which can represent the input signal
to the chain of p amplifiers and loudspeakers in the speech production system. The measurement
sound ym (t) is, by definition, considered to be a superposition of the response y (t) = Hu (t) and
the unmodeled contribution e (t), which are objects to be modified and controlled. Of course, as a
practical matter, the requirements for good results include modeling and system design such that
the magnitude | e (t) | is not greater than the magnitude | y (t) | in critical frequency bands It is
to The main purpose is to change the dynamics of the speech production system represented by
Equation 11 in certain reference dynamics. For this purpose, we introduce a criterion matrix D.
Where w (t) is a set of raw or recorded sounds, or an artificial including the test signal used in
the filter design. <Img class = "EMIRef" id = "197718648-000004" Is an r-dimensional vector
that also includes the digital audio signal generated by. The elements of the vector w (t) may
represent, for example, a channel of digitally recorded sound or a sampled and digitized analog
source. D in Equation 12 is a transfer function matrix of dimension m × r assumed to be known.
The linear system D is variable in design and generally represents the reference dynamics of the
vector y (t) in Eq. An example of a possible design objective is the complete inversion of the
dynamics and the separation of the channels. In the case of r = m, the matrix D is set equal to a
square diagonal matrix with the d-step delay operator as the diagonal component.
The nominal response of y (t) is defined as being a delayed version of the original sound vector w
(t), w (t) With a delay d of equal sampling period for all elements of. In more complex designs, in
addition to introducing delays, reference dynamics may be added to the speech production
system in the form of a stable filter. With such a design of D, it is also possible to add new sound
features to the system. For example, it is possible to obtain excellent sound quality in low quality
audio equipment. For example, a more complex design may be important to emulate the specific
sound of one type of speech production system. The desired bulk delay (d) introduced through
design matrix D is an important parameter and affects the achievable capacity. The causality
compensation filter achieves better compensation than this delay is acceptable. The precomp is
generally obtained by a precomp filter indicated by R, and the speech reproduction system
equation based on w (t) of the signal. 11. Generate an input signal vector u (t) for. In the prior art,
the superior tendency of digital speech precomp is to generate an input signal vector u (t) to the
speech reproduction system equation 11, whereby its compensated output y ( It is said that t)
approximates the reference vector yref (t) in a certain sense. This objective can be achieved by
the signal u (t) in equation 11 being generated by a linear precompact filter R. The filter R
operates on the signal w (t) such that the element consists of a stable linear causal dynamic filter
p × r matrix and the signal y (t) approximates yref (t). <Img class = "EMIRef" id = "197718648000007" /> Within the system general range, the condition for accurate compensation is that R
multiplies the dynamic model H (the right side of D It is equal to the right inverse matrix with
stable causality in).
Where H <−R> represents the right inverse matrix of the transfer function matrix of the model.
<Img class = “EMIRef” id = “197718648-000008” /> Such right-handed inverse matrix by
definition has the property HH <-R> = Im (an identity matrix of size m × m). したがって、
HR=HH<−RD>=Dである。 Unfortunately, models of speech systems often do not have
accurate stability causal right inverse matrices. However, given that the bulk delay d in the range
of D (the smallest delay due to any element of D) can be increased, the least squares
approximation error | y (t) achieved by the stable causality compensation filter yref (t) | 2
indicates that if the normal rank of H (the rank other than the transfer function matrix in a zero
matrix system) is equal to the number of elements in m × y (t), it disappears as a delay d it can.
The delay d is determined by the designer. The designer can thereby control the degree of
approximation. In order to be able to work well, the system described by H needs to have at least
as many separate inputs as outputs, ie p ≧ m. On the other hand, the rank of H can never be as
large as m. In the simplest case, y (t), u (t) and w (t) are all scalar time series, with m = p = r = 1 in
scalar model and scalar reference dynamics. モデルHは。 It can compensate the amplifier and
loudspeaker chain. In the prior art and the paper, the most promising way of solving this kind of
approximation problem is focused on representing H and R by FIR filters, and then to minimize
the scalar type criterion Using the least squares technique, we penalize the sum of the mean
squared differences of the elements of y (t) and yref (t). <Img class = "EMIRef" id = "197718648000009" /> In the following, () <T> denotes a transposed matrix of vectors, and E () is a related
statistic of the signal to be contained Represents the average on the physical characteristics. Such
a least squares design can achieve on-line recursive minimization of Eq. 17 by applying, for
example, the LMS algorithm or the filtering xLMS algorithm to the measured signals ym (t) and w
(t) (Patent Documents 12 and 13).
Also, this design can be performed off-line by solving the Wiener optimization problem for a
fixed degree of FIR filter. This is equivalent to solving a set of simultaneous binary linear
equations and a Wiener-Hopf equation, including correlation evaluation. The minimization of
Equation 17 should take into account not only the amplitude response but also the phase
response of the system. This approach is better than a method that only takes into account the
amplitude response (see, eg, reference 1). The drawback with the use of FIR filters is that filters
with very large coefficients may have to be used. For this reason, the present invention focuses
on the adjustment of IIR filters, which generally requires less coefficients. Regardless of the use
of FIR or IIR filters, careful analysis by the inventor has shown that all prior art designs based on
the minimization of the least squares criterion shown in Equation 17 suffer from a further
important drawback. The compensation filter based on the minimization shown in Equation 17
obtains extreme characteristics at the highest and lowest frequencies. In the case of the scalar,
this results from the transfer function H, which often has a decreasing gain at the highest and
lowest frequencies in the range of speech, and results in a compensator R with high gain at these
frequencies. Such a compensator has a long and oscillating impulse response (see FIG. 4). It
requires operational adjustment and implementation. This is a potential problem not only at
extremely high or low frequencies, but also across all frequencies where excessive compensation
is required, if the criteria in Eq. 17 are minimized. Furthermore, a compensation filter R with too
high gain at certain frequencies can cause non-linear distortion and have deleterious effects. In
the worst case, too high a gain input can damage the audio equipment. It will be appreciated that
better control over the range and amount of compensation performed at different frequencies
and different subchannels needs to be achieved than with Eq. In the design of the precomp filter
for the audio installation according to the invention, the filter additively comprises two
components, a fixed and non-zero filter element and an adjustable compensator component
determined by optimization. It proves to be useful. Fixed filter components are typically
configured by the filter designer or set to a default configuration.
The adjustable compensator approximates the pre-compact filter to a fixed and non-zero filter
component while the component is determined by optimizing the criterion function based on the
given weighting, and pre-compensating for the reference system response Approximate the
model response that has been sunk. Although not required, this weighting is preferably
frequency dependent and / or channel dependent. This is illustrated below. In order to further
understand the basic concepts of the present invention, the design of such weighting based
precomp filters is described. For example, the compensation can be realized by additionally
changing the signal path as m (t) = Cw (t). It is usually the direct feedthrough and delay of w (t) of
the signal. Where g is a suitable delay, and C is a matrix of FIR or IIR filters. <Img class =
"EMIRef" id = "197718648-000010" /> In Equation 18, u (t) and w (t) are assumed to have equal
dimensions (m = r). Using the standard backward shift operator notation: <img class = "EMIRef" id
= "197718648-000011" />, in Eq. 14, the compensator matrix is such a form for design It is
considered to have The design of the compensator component C is preferably such that the
additional transform signal m (t) = Cw (t) Based on the minimization of a reference transform that
includes frequency weighting periods that penalize magnitude. The penalty period may be of any
type of criteria used for filter optimization. In particular, when replacing the second-order
criterion function shown in Equation 17, the following equation is obtained: <img class =
"EMIRef" id = "197718648-000013" /> Here, W is a first weighting function, and V is an
additional optional weighting function. The matrix W is preferably a square m × m matrix and
includes stable linear IIR filters and represents a set of design variables.
Furthermore, the additional weighting function V is preferably a square p × p matrix, includes a
stable linear IIR filter, and can be used as another set of design variables. In a particular
embodiment of the invention, the weighting represented by the transfer function matrix W acts
as a frequency dependent penalty on the compensation signal m (t) = Cw (t). The effect of
weighting by W is best understood in the frequency domain by using the Z-transformed
representation of the signal and system. The minimization of Equation 21 occurs at a
compensation period C (z) where the W (z) criterion has a small gain at a relatively large
frequency z. This is because the last period of equation 21 governs J on the one hand. In such a
frequency band, C (z) w (z) is small in Equation 18, and the uncompensated system properties
remain unchanged, except for the g sample delay. On the other hand, at frequencies z where the
W (z) criterion is negligible, the first period of the criterion of Equation 21 is most important. In
the case of V = i, in these frequency bands, y (z) to yref (z) = D (z) w (z) (... indicates that they are
roughly equal. same as below. ). This adjustment is to minimize the contribution of the values of
the first period of Equation 21. For example, the weighting function represented by W may be
implemented as a low pass filter at a given break frequency, in parallel with a high pass filter
with a given limiting frequency. By properly choosing the break and limit frequencies, the
compensation performed by the precomp filter can be customized to the particular application.
Of course, the weighting W can be realized in any form. The frequency-selective weighting by
matrix V can be used for various purposes, such as: * Can use known features of the human ear
for weighting of perception. The elimination of compensation errors in the frequency band to
which we are more sensitive is emphasized. * H modeling errors can be used for low weighting
with capacity deviations in large frequency bands, and the optimization will not focus on
unreliable frequency bands. Weighting the errors achieved in different places in space, ie
different components of the vector y (t), can also be used. This can be achieved by setting V equal
to the diagonal transfer function matrix and by using different filters as the diagonal elements of
Although design model H describes all relevant frequency ranges, the use of frequency
dependent weighting allows for different adjustments in different frequency bands. In this way, it
is possible to avoid the solution of dividing the entire frequency band into secondary bands and
compensating for each band separately. More complex side band solutions (eg, used in graphic
equalizers) are known to cause problems with phase response distortion. It should also be noted
that W forms the basis of a weighting filter in the case of multiple channels. It is possible to use a
diagonal matrix with each different diagonal element to individually adjust the compensation
performed on each input channel for a particular loudspeaker characteristic. This type of channel
dependent weighting uses frequency independent weighting or frequency dependent weighting
for each channel, and independently enables different types of compensation in different paths of
the multi-channel system. It can be implemented. The direct feedthrough (or bypass) delay g of
Equation 18 is yet another design variable. In the case of scalars, the appropriate choice is (m = p
= r = 1) if d> = k is set to g = d−k. Here, k is a bulk delay of H and d is a bulk delay of D. Thus,
the total net delay due to the compensated system is g + k = d in almost all frequency bands. If u
(t) ~w (t−g) is obtained in a band that is considerably penalized by W, then the total delay of the
compensation model HR is g + k. In the band where W is insignificant, it is HR ~ ~D, and D is
assigned a delay d in advance. For multi-channel compensators, different feedthrough delays as
well as different bulk delays of D may be required in different paths. Such channel dependent
delay helps to generate virtual sound, i.e. to make the sound appear to originate from directions
other than the loudspeaker. To include such and other variables of the compensation problem,
and to handle cases where the number of signals in w (t) is different from the number of signals
in y (t), Equation 18 is generalized as r ≠ m Ru. Where F is an arbitrary m × r matrix of a stable
linear dynamic system. <Img class = "EMIRef" id = "197718648-000014" />
This matrix is assumed to be known and not modified by the optimization. The special case that F
is zero corresponds to using a penalty on the compensator output u (t) and is identical to m (t).
This special case was discussed in the prior art on a quadratic basis with a special weight
selection with V = 1 and W equal to the frequency independent weight as a special case of scalar
type systems (see reference 4). Such optimized feed-forward regulators have been designed for
process control (see references 5, 6). This type of design has been found to be unsuitable for
speech pre-comp and is therefore excluded from proposing solutions. The large penalty W
cancels the magnitude of the total signal vector u (t) which is itself a large distortion of the
existing system characteristics, for F = 0. The main purpose of the proposed compensator design
is to introduce a penalty instead. It may leave the natural response of the invariant system, here
to get large W and F = q <-g> I. An important element in the proposed design is to assume that
the compensator (Eq. 14) can be additively decomposed into two parts for design purposes.
Where F is a fixed non-zero filter component, and C is the goal of the optimization. <Img class =
"EMIRef" id = "197718648-000015" /> Equation 18, which is a special case of Equation 23,
corresponds to F = q <-g> I for r = m. The fixed non-zero filter component F may be a simple
bypass component with such selectable delays. However, it does not prevent F from being
configured with one or more additional fixed filtering components. In general, the design
principle proposed for obtaining C at compensator 23 is to optimize the criteria that include two
objective weights. i) as small as possible the deviation between the overall precompensator filter
R and the predetermined dynamic non-zero filter component F, and ii) the compensated design
model HR and the predetermined dynamic reference system D Make the deviation between them
as small as possible. In particular, when this weighting is frequency dependent and / or
dependent on the input channel, an efficient tool for automated computer aided filter design is
obtained, which may be at different frequency bands and / or multiple channels. Provides control
over the amount of compensation performed on different subchannels in the design.
The precomp filter of the present invention is mainly implemented as a digital filter or a set of
digital filters in a multi-channel system. The filter and model can represent any operator and can
perform transformations suitable for linear systems. For example, it can be represented by a
delay operator form, a Z conversion representation, a delta operator representation, a series
representation of a function or a frequency representation (see reference 7). The degree of
approximation (nearness) can be measured with any reference matrix of linear invariant dynamic
systems, such as quadratic reference 21, frequency weighted H∞-references, L1-criteria (see
references 8 and 9). In order to better understand the advantages of the present invention, a
comparison of the ability of the precomp filter designed according to the present invention and
the precomp filter designed according to the prior art is made. In this example, the precomp filter
is applied to one loudspeaker coupled to an amplifier. The amplitude response and deviation of
the phase response of the modeled speech chain is shown in FIGS. 2A and 2B, respectively. A
typical impulse response is also shown in FIG. The sampling frequency is 44.1 kHz. The design
model has a zero bulk delay k, but its impulse response is shifted to the right in FIG. 3 to make a
simple comparison with the compensation response. As a desired criterion in Equation 12, y ref
(t) = w (t-d) was used with d = 300 samples. As can be seen from FIG. 2A, the amplitude response
of the uncompensated experimental loudspeaker and amplifier model is by no means ideal: there
is ripple in the mid frequency band and low power at low and high frequencies. First, this
experimental model was compensated by minimizing (Eq. 21) with the feasible (stable and
causal) IIR compensator (Eq. 18) according to the present invention. Use the multi-term Wiener
design described in more detail in Section 2. Complete inversion of the entire audio range from
20 Hz to 20000 Hz requires extreme amplification at the lowest and highest frequencies in FIG.
2A. If the entire speech range is inverted, a compensation signal of too high power may be
generated, especially for the highest and lowest frequencies. Such strong signals may damage the
audio equipment. Thus, instead, the loudspeaker dynamics (up to delay d = g = 300) are
completely inverted in the frequency range from 80 Hz to 15 kHz.
Also, the amplification should be less than 20 dB outside this range. The weighting W in this
example (equation 21) uses a low pass filter with a 30 Hz break frequency, which is used in
parallel with the high pass filter with a 17 kHz limiting frequency (see FIG. 8). The impulse
response of the designed IIR precomp filter is shown in FIG. The compensated amplitude
response and deviation of the phase response are shown in FIGS. 6A and 6B, respectively. As
seen in FIG. 6A, the intermediate frequency ripple in FIG. 2A is removed and the amplitude
response in the compensated frequency range (80 Hz to 15 kHz) is close to the required flat
response (amplitude response = 0 dB) Do. Also, the deviation of the phase response of the
compensated model system (FIG. 6B) was significantly improved compared to the
uncompensated deviation of the phase response in FIG. 2B. The compensated impulse response
(shown by FIG. 7) is close to the ideal Dirac pulse response yref (t) = w (t-300). The remaining
small ripple near the main peak is due to the fact that it has limited the amount of complement at
the lowest and highest frequencies. This ripple can be eliminated by designing with W = 0 at the
expense of designing the precomp filter at very high gain at the lowest and highest frequencies
(see FIG. 9). These results are then compared to the FIR filter precompensator designed by the
minimization of the least squares criterion (Eq. 17), using the LMS algorithm as the ideal, with
appropriately adjusted step lengths . Such compensators have a long, oscillating impulse
response that requires computational tuning. This is a potential problem for all frequencies
where excessive compensation is required, not only at the highest and lowest frequencies but
also when the criterion (Eq. 17) is to be minimized. The amplitude response and associated phase
response of a prior art compensated system are shown in FIGS. 10A and 10B, respectively. The
amplitude response of this compensated system shows much more oscillations for the
intermediate frequencies and in particular for the highest frequency compared to the system
compensated with the filter according to the invention. Thus, the design of the present invention
is shorter and provides a better compensation filter and provides more accurate inversion within
the range of frequencies for which compensation is required. <Section 2: Scalar-Type
Compensator Designed as Causal WIENER Filter> A precompact filter design method in which the
scalar type filter is designed as a causal WIENER filter will be described below with reference to
FIG. .
As an embodiment of the present invention, the problem of precompensating one audio chain
(amplifiers, cables, loudspeakers and if necessary room acoustics) is considered. The scalar model
H may represent an average over the measured dynamics at many points as compared to the
loudspeaker, such that the volume of space where good compensation is achieved is large. Some
room acoustic responses are ignored so that only the loudspeaker chain is compensated. In this
case, linear systems and models are all assumed to be invariant. They are shown using a discrete
time backward shift operator denoted q <-1>. The signal s (t) is one sample backward shift by this
operator q-1s (t) = s (t-1). Similarly, the forward shift operator is denoted by q, so qs (t) = s (t + 1)
(see, eg, reference 10). The scalar design model Equation 11 is represented by a linear invariant
difference equation with fixed coefficients. <Img class = "EMIRef" id = "197718648-000016" />
Assuming that b ≠ 0, k before k (t) input affects output y (t), There is a delay of one sample. This
delay (k) may, for example, represent an acoustic transport delay. This is called the model bulk
delay. The maximum delay n, h may be hundreds or thousands of samples in some models of
audio systems. Move all items related to y to the left side. In the shift operator representation,
model 24 is equal to: <Img class = "EMIRef" id = "197718648-000017" /> polynomial A (q <-1>)
= (1 + a1q <-1> + a2q <-2> +... + Anq <-n> Discrete time dynamic model (Equation 24) is shown in
a more compact representation by introducing B) and B (q <−1>) = (bo + b1q <−1> +... + Bhq
<−h>) obtain. The polynomial A (q <-1>) is said to be monic because its principal coefficient is
one. <Img class = "EMIRef" id = "197718648-000018" />
In particular, A (q <-1>) = 1 in the FIR model. In general, the recursion at the old output y (t-j)
represented by the filter A (q <-1>) gives the model an infinite impulse response. Also, the IIR
filter shown in Equation 26 is a logical filter. This is because those conversion operators can be
represented by the ratio of polynomials q <-1>. <Img class = "EMIRef" id = "197718648-000019"
/> [0060] All complex IIR systems, models and filters are assumed to be stable below. When the
complex variable z is substituted for the operator q, the stability criterion means that the
equation A (z <-1>) = 0 with solution only if it has size | z | <1. In other words, the complex
function A (z <-1>) must have all zero matrices within the unit circle in the complex plane. The
hypothetical second order statistics (characteristics of the spectrum) of w (t) of the signal to be
compensated can be expressed in a stable and stable reversible Auto-Regressive Moving Average
(ARMA) model. <Img class = "EMIRef" id = "197718648-000020" /> where v (t) is white noise.
Also, the polynomials H (z <-1>) and G (z <-1>) have all zero matrices at | z | <1 since both
highest order coefficients are equal to 1, ie stable ing. A design model (Equation 12) that shows
the required response for y (t) is given by the stable difference equation. Where the polynomial N
(q-1) is monic and the principal polynomial of the main polynomial at D (q <-1>) Since the
coefficients are assumed to be non-zero, d represents the required bulk delay. The compensator
structure used is as shown in Equation 23, the fixed filter F is (polynomial) F (q <−1>), and the
bypass delay g is d − assuming that d ≧ k. Equal to k.
Temporarily made sense in the section above. Thus, by minimizing the second-order criterion (Eq.
21), a stable discrete-time scalar logic filter C ( q <-1>) is optimized. Here, Wm (t) is a scalarstabilized dynamic system based on the output f (t), and is regarded as V = 1 for simplification,
and is expressed by the differential equation as follows. <Img class = "EMIRef" id = "197718648000023" /> Both polynomials V (z <-1>) and W (z <-1>) are design variables. They are restricted
to have all zero matrices at | z | <1. Thus, the criterion (Equation 21) can be expressed as follows.
<Img class = “EMIRef” id = “197718648-000024” /> The optimization of the solution is
described below. The model described above and shown in FIG. 11 and the filter polynomials V,
W, G, H, D, N, B, A and the delays k, d are specified numerically. The stable and causal IIR filter C
(q <−1>) of Equation 30 minimized by the criterion (Equation 32) is specified by the difference
equation. <Img class = "EMIRef" id = "197718648-000025" /> Here, the monic polynomial (q <1>) has all zero matrices at | z | <1. That is, together with the scalar r, the unique and stable
highest-order coefficient of the multiplicative spectrum factorization equation is given as a
solution equal to one. Here, Q (q <−1>) in polynomial 33 is polynomial with a non-causal FIR
filter L * (q). It is given by the unique solution of the linear scalar Diophantine equation.
<Img class = "EMIRef" id = "197718648-000027" /> In the above, the polynomial in the forward
shift operator is a noncausal operator that shifts the signal forward in time. Represents These are
indicated with an asterisk * as a subscript. For a polynomial P (q <-1>) = (p0 + p1q <-1> + p2q <2> +... + Pnpq <-np>) with real-valued coefficients, let P * (q) = It is defined as (p0 + p1q + p2q
<2> +... + Pnpq <np>). While it is assumed that N (q <-1>) and G (q <-1>) have all zero matrices at
| z | <1 for formulation of the problem, while β (q <− The filter (Eq. 33) is guaranteed to be
stable since 1>) has zero matrix only at | z | <1. Compensators are causal. The reason is that the
related filter has only the backward shift operator as an argument, and has the non-zero main
coefficient due to the fact that βGN is monic in Eq. This means that m (t) and its output u (t) is a
function of the future value of w (t) at time t. The optimal filter structure (Equation 33) and the
corresponding design equation (Equation 34, 35) can be derived by the orthogonal principle
equation (see, eg, references 6, 10, 11, 16). All eligible alternative filters are considered if the
alternative compensator can only achieve a criterion value lower than the criterion value
obtained by equation 33. The multinomial spectrum factorization equation (Equation 34) always
has a stable solution. When the compound variable number z is substituted for the operator q,
the right side of Formula 34 can be considered as a polynomial with zero matrices arranged
symmetrically on the inside and outside of the unit circle | z | = 1. The zero matrix can not be
exactly located on the unit circle because of the above-mentioned filter and model stability
assumptions. The solution of Equation 34 corresponds to a set of unique elements including all
zero matrices inside the unit circle forming the polynomial β (q <-1>). The scalar r is just a
normalization element that makes up the β (q <-1>) monic expression. The polynomial
Diophantine equation (Equation 35) can be easily converted to a linear system, and can be solved
for polynomial coefficients Q (q <−1>) and L * (q).
These equations are formed by multiplying the equal power coefficients in the equation q on
both sides of Equation 35. General theory for solving multinomial Diophantine equations (see
reference 12), Equation 35 can be guaranteed to have a unique solution. This is because the right
side of the polynomials β * (z) and A (z <-1>) N (z <-1>) H (z <1>) z can never have a common
factor. This is a conjugate polynomial of β * (z) being β (z <-1>), and A (z <-1>), N (z <-1>) and H
(z <1>) are According to the design conditions, a zero matrix is provided inside | z | = 1, while all
of the zero matrix is outside | z | = 1. As described above, the design problem with a defined
condition can always be solved, and the solution is embodied by the compensation filter
(Equations 30, 33) and the design equation (Equations 34, 35). Linear invariant filters that
minimize quadratic criteria based on quadratic (spectral) signal models are literally referred to as
Wiener filters (see, for example, reference 13). Not only in the area of speech precomp, but also
in the design of the Wiener filter and the linear quadratic design in general, the compensator
design equation of the filter (Equation 30) by minimization of the criterion (Equation 32) shows
new results. <Section 3: Multivariate Compensator Implemented in State-Space Form Designed by
Linear Quadratic Optimization etc.> The form and design of the polynomials in the above section
are the polynomials described in reference 14 The matrix representation can be generalized to
MIMO (multiple input, multiple output) filters and models. Also, MIMO design can be performed
by linear quadratic Gaussian distribution (LQG) optimization based on the state space model and
its design. This is described below. For a general introduction of LQG designs based on state
space methods, see, for example, document [28]. In the following, the conventional notation of
dynamic systems in the state theory field is used to describe the multi-channel function of the
precomp filter according to the invention. In the following, a matrix whose elements are realvalued constants (not a filter) is indicated in the text by subscripting the character that is the
symbol in front of the character that is the symbol, and the character that is the symbol in the
mathematical expression In bold and underlined. The vector-ARMA model of w (t) is introduced
as a linear discrete time-invariant state space model, along with the state vector x1 (t) of
appropriate dimension.
Where w (t) is a column vector with dimension r, as in the case of section 1. <img class =
"EMIRef" id = "197718648-000028" /> The r-dimensional vector v (t) represents white noise
with a known covariance matrix -R1. The ARMA model (Equation 36) is considered stable and
stable reversible. In Equation 36, −D1 is assumed to be a reversible r × r matrix, and is
generally set equal to the unit matrix. Assuming that w (t) is white noise, the dimensions of x1 (t)
are zero and w (t) =-D1v (t). A stable linear design model H (Equation 11), which describes the
audio system to be compensated, is implemented in state space form with a state vector x 2 (t).
Here, the vector y (t) is a dimension m, and u (t) is a dimension p. Bulk delay is assumed to occur
in the state delay structure. Thus, a larger delay increases the dimension of the state vector x2 (t).
A stable demand system (Equation 12) is also realized in state space form with a state vector x3
(t). <Img class = "EMIRef" id = "197718648-00300" /> Here, the bulk delay d is incorporated into
the state delay structure. Using the compensator filter structure (Eq. 23), a stable preset linear
filter F is implemented in state space form with a state vector x 4 (t). <Img class = "EMIRef" id =
"197718648-000031" /> In Eq. 39, the additional signal m (t) is the criterion here as V = I for
simplicity. 21) optimize based on Nevertheless, the stable input penalty filter W in the criterion is
implemented as another filter in state space form with an output signal whose vector is shown as
f (t). <Img class = "EMIRef" id = "197718648-000032" /> Here, the quadratic criterion to be
minimized (Equation 21) is given as follows.
<Img class = "EMIRef" id = "197718648-000033" /> Then, the state vector of the entire system is
defined as follows. <Img class = "EMIRef" id = "197718648-000034" /> This state update
equations 36 to 40 are also combined into a single model. <Img class = "EMIRef" id =
"197718648-000035" /> Here, the state transition matrix -F of the joint model and the input
matrix -G, -H are secondary models (Equation 36 to Equations 36) It is easily obtained from 40).
The criterion (Eq. 41) can then be expressed in the form of a criterion with infinite control
horizontals and penalties in the selected state. Add a penalty with penalty matrix -R on the
quadratic of m (t) as a normalization period. <Img class = "EMIRef" id = "197718648-000036" />
Here, <img class = "EMIRef" id = "197718648-000037" />. If x (t) is known, then linear state
feedback: <img class = "EMIRef" id = "197718648-000038" /> minimizes the infinite horizontal
criterion (Eq. 43) It may be possible. The optimal controller gain matrix is given by Where S is a
symmetric, semi-positive definite matrix that solves the algebraic matrix Riccati equation. <Img
class = "EMIRef" id = "197718648-000039" <Img class = "EMIRef" id = "197718648-000040" />
The overall system is by definition detectable and stabilizable, as all complex systems are stable.
This guarantees the existence of a solution to this linear secondary state feedback control
problem. This solution corresponds to the semi-positive definite matrix solution -S for Eq. When R is set to a positive positive definite matrix, there is always a p × p matrix appearing as an
inverse matrix in Eq. 47 and Eq. If the state vector is not known, it can be evaluated by the
observer. The separation principle of linear quadratic optimal control theory is that if this
observer is designed as a quadratic optimized linear observer (Kalman Estimator), it is minimized
using only the measurement signal (formula 44) It states that it can be obtained by joint optimal
design. Such designs are known as linear second-order Gaussian (LQG) designs or H2-optimal
designs. In the formulation of the particular problem considered here, the observer in the optimal
state is simple in design. The stable subsystems (Eqs. 38 to 40) can be derived without noise with
the measurement signal alone, which form part of the compensator and problem formulation.
Their state is thus known. The output of the model (Equation 37) is not directly measurable. This
is because the design is not designed to solve by feedforward, and does not use feedback from
the voice measurement ym (t). The most qualified observer for x2 (t) is a copy of Equation 37
driven by u (t) of the known signal, providing a state estimate x2 (t | t-1). Since -D1 is assumed to
be reversible in model 36, the noise input v (t) can be estimated as follows: Therefore, the state
estimate for x1 (t) can be updated by the following equation: <img class = "EMIRef" id =
"197718648-000041" /> <Img class = "EMIRef" id = "197718648-000042" /> [0084] This
recursion is stable since it is assumed that the ARMA model 36 is stable and invertible. Assuming
that w (t) is white noise, Equation 50 is of course excessive. The complete solution is thus given
by Eqs. 50, 37, 38, 40 for evaluating the state and Eq. 39 representing the precompensator, and
m (t) occurs as follows.
<Img class = "EMIRef" id = "197718648-000043" /> Here, <img class = "EMIRef" id =
"197718648-000044" /> The compensator (Equation 39) , 51) becomes <img class = "EMIRef" id
= "197718648-000045" />, and the r input w (t) and the p output u (t) constitute an IIR filter.
The gain matrix-L is optimized by solving equation 48 for -S, which is one of the many existing
solvers of Riccati equation. Then use equation 47. Section 4: Nonlinear Models and
Compensators The design principles introduced in section 1 can be generalized to the speech
precompilation problem. In that case, the design model may be non-linear and / or the required
compensator may have a non-linear structure. The simplest example would be a linear system
and compensator in series with a non-linear static element (eg, a limiter). Elements like these are
in fact always present in real systems but are ignored in linear design and optimization. Other
conceivable non-linear models and filter structures include model structures consisting of
Volterra and Wiener models, neural networks, series expansions of functions and non-linear
physical models of acoustic elements. Define a set of vectors of delayed signals. <Img class =
"EMIRef" id = "197718648-000046" /> A non-linear and variable dynamic model corresponding
to Equation 11 can be expressed as follows. Where h () denotes a dynamic operator that may be
non-linear and variable. <Img class = "EMIRef" id = "197718648-000047" /> Similarly, a nonlinear demand response model that generalizes the structure (Equation 12) can be expressed as:
<Img class = "EMIRef" id = "197718648-000048" /> where d () denotes a dynamic operator that
may be non-linear and variable. An important feature of the present invention (including the nonlinear case) is the additional degradation of the precompensator. For a compensator that will be
non-linear and variable, this is expressed as: <Img class = "EMIRef" id = "197718648-000049" />
where r (), f (), and c () are non-linear, time-dependent stable dynamic operators Indicates The
operator f is preset and is also not zero while c is to be adjusted by optimization. Preferably, a
parameterization of c, for example c = 0, is allowed for the setting of the parameters, such that in
such a case a nominal response r = f is obtained. Also for nonlinear problems, the optimization
criterion is weighted between the closeness of r to f (smallness of m (t)) and the closeness of the
compensation output y (t) to yref (t) Must be included. If this weighting is frequency dependent,
it has to be represented by a linear and stable dynamic weighting matrix V, W as in the linear
case. This is because the frequency response is preserved in a meaningful form only by the linear
system. The criterion corresponding to Equation 21 depends on the amplitude of the input signal
for non-linear systems. A scalar quadratic criterion that weights the response for a given
sequence of decision input signals w (t) may be defined and minimized. Possible and appropriate
criteria can be expressed as: Where imgt () denotes the sum over a particular test signal
sequence w (t) with an appropriate amplitude range. <Img class = "EMIRef" id = "197718648000050" Minimization of the free parameters of c () in Equation 57 could be performed by a
mathematical search routine for nonlinear models and filters. Section 5: The Most Primitive Face
In general, the design equations are solved on separate computer systems to create the filter
parameters of the precomp filter.
The computed filter parameters are typically downloaded to digital filters, such as digital signal
processing systems or similar computer systems that perform the actual filtering. The filter
design mechanism according to the invention is preferably implemented as software in the form
of program modules, functions or equations. The software may be written in any type of
computer language (eg, a specialized language for C, C ++ or Digital Signal Processor (DSP)). In
practice, the relevant steps, functions and actions of the present invention are mapped to a
computer program and executed by the computer system to perform the operations associated
with the design of the precomp filter. In the case of personal computer (PC) based systems, the
computer program used to design the audio precomp filter is usually a code that is read by a PC
readable medium, such as a CD for distribution to users and filter designers. Turn The user or the
filter designer may load the program from these media into the computer system for subsequent
execution. FIG. 12 is a schematic block diagram showing an example of a computer system
suitable for the function of the filter design algorithm according to the present invention. System
100 may be implemented on any known computer system, including a personal computer (PC),
mainframe computer, multiprocessor system, network PC, digital signal processor (DSP), and the
like. System 100 basically includes a central processing unit (CPU) or digital signal processor
(DSP) core 10, a system memory 20 and a system bus 30 interconnecting various system
components. The system memory 20 includes read only memory (ROM) 22 and random access
memory (RAM) 24. Additionally, system 100 includes peripheral storage device 40, which is
typically controlled by one or more drivers. The peripheral storage device 40 is a non-volatile
storage device of data and program information of, for example, a hard disk, a magnetic disk, an
optical disk, a flexible disk, a digital video disk or a memory card. Each peripheral storage device
40 is associated with a drive for controlling the storage device as well as a drive interface (not
shown) for connecting to the system bus 30.
The filter design program implementing the design algorithm according to the present invention,
together with other related program modules, is stored in the peripheral storage device 40 and
loaded into the RAM 22 of the system memory 20 for execution by the CPU 10. The filter design
program operates on the filter parameters of the precomp filter given the associated input data
(e.g. model representation, fixed filter components, configured weighting and reference system
representation). The determined filter parameters are transferred from the RAM 24 in the system
memory 20 to the precomp filter system 200, typically via the I / O interface 70 of the system
100. The precomp filter system 200 comprises a signal processor (DSP) or similar central
processing unit (CPU) 202, having one or more memory modules 204 for storing filter
parameters and samples of required delay signals. Is desirable. The memory 204 also typically
includes a filter program, which when executed by the processor 202 performs the actual
filtering based on the filter parameters. Instead of transmitting the filter parameters computed
directly via the I / O system 70 to the precomp filter system 200, a memory for later distributing
the filter parameters to the precomp filter system It is also possible to store it on the card or
peripheral storage disk 40 and save the data away from the filter design system 100. A known
microphone unit or similar recording facility 80 is connected to the computer system 100, for
example via an analog to digital (A / D) converter 80, in order to be able to measure the sound
output by the audio facility You may Based on measurements of known test signals obtained by
the microphone unit 80, the system 100 can refine the model of the audio system using an
application program loaded into the system memory 20. The measurements can also be used to
assess the performance of the precomp filter and the combined system of the audio equipment. If
the designer is not satisfied with the resulting design, the designer may begin a new optimization
of the precomp filter based on the modified set of design parameters.
Additionally, system 100 may have a user interface 50 between the filter designer and the user.
Several different user interface scenarios are possible. For example, the filter designer may
decide to use a particular set of customized design parameters (e.g., a particular fixed filter
element or weighting in the operation of the filter parameters of the filter system 200). The filter
designer then defines through the user interface 50 fixed filter components and associated
design parameters such as weighting. The designer can choose between different pre-set fixed
filter compensation and / or weighting sets. These sets may be designed to produce different
features of the different audio systems, listening environment and / or sounds. In such cases,
preset options are typically stored in ambient storage 40 and loaded into system memory during
execution of the filter design program. The filter designer chooses fixed non-zero filter
components and / or maximum weighting by testing some of the preset options and / or
changing the parameters in the preset options The audio system and listening environment can
be adapted. Alternatively, the filter design program may automatically choose fixed and default
non-zero filter components and weights, preferably based on the audio equipment for which the
precomp filter is to be used. In addition to fixed non-zero filter components and frequency
dependent and / or channel dependent weighting, filter designers can also define a reference
system by using the user interface 50. For example, the reference system delay can be selected
by the user and provided as a default delay. It is also possible to select and introduce more
advanced special effects to the reference system. Such special effects may include enabling a
cinema-like audio reproduction in a compact stereo system. Instead of determining the system
model based on the measurements of the microphones, it is also possible for the filter designer to
select a model of the speech system from a set of different pre-set system models.
Such selection is preferably based on the particular audio equipment using the precomp filter
according to the invention. As another embodiment, an example of construction in which the
filter design is independently executed with or without the participation of the peripheral users
will be described below. A typical system includes an input / output monitoring program, system
identification software and filter design software. The input / output monitoring program first
generates a test signal to measure the acoustic response produced by the voice system. Based on
the test signal and the obtained measurements, system identification software determines a
model of the audio system. The I / O monitoring program gathers and / or generates the
necessary design parameters and sends these design parameters to the filter design program
which computes the precomp filter parameters. The input / output monitoring program can
optionally evaluate the design capability with the measured signal and, if necessary, determine a
new set of filter parameters based on the changed set of design parameters for the filter design
program. It can be done. Repeat this procedure until a satisfactory result is obtained. Then
download the final set of filter parameters to the precomp filter system. It is also possible to
adjust the filter parameters of the precomp filter appropriately, without using a fixed set of filter
parameters. Audio conditions may change while using the filter in an audio system. For example,
the position of the loudspeaker or the position of an object such as furniture in the listening
environment may change. These can affect the acoustics of the room. Also, some equipment in
the audio system may be replaced with other equipment to change the characteristics of the
overall audio system. In such a case, one or more microphone units or similar recording
equipment will perform continuous or intermittent measurement of the sound from the audio
system at one or several locations in the listening environment Will. The recorded audio data is
provided to a filter design system, such as the system 100 of FIG. The system computes new
audio system models and adjusts filter parameters to suit newer audio situations. Of course, the
invention is not limited to the example of FIG.
As an option, the design of the precomp filter and the actual implementation of the filter may
both be done in one and the same computer system 100 or 200. This means that the filter design
program and the filtering program run on the same DSP or microprocessor system. A speech
production system 300 employing the precomp filter system 200 according to the present
invention is schematically illustrated in FIG. The audio signal w (t) from the sound source is
transferred to the precomp filter system 200. This transfer occurs through the conventional I / O
interface 210. If the audio signal w (t) is an analog signal, such as, for example, a sound from an
LP record, an analog recording cassette tape or other analog source, the signal is first digitized in
the A / D converter 210 before entering the filter 200 Be done. Digital audio signals from CDs,
DAT tapes, DVDs, mini-discs, etc. can be transferred directly to, for example, the filter 200
without conversion. The digital or digitized input signal w (t) is then precompensated by the
precomp filter 200, basically taking into account the effects of the subsequent voice system
equipment. The compensation of the digital speech signal varies according to the frequency
dependent and / or channel dependent penalty period and penalizes the compensation part of
the filter system. The resulting compensated signal u (t) is then sent to the DA converter 240 via
the further I / O device 230 and the digital compensation signal u (t) is converted into the
corresponding analog signal Ru. This analog signal then enters amplifier 250 and loudspeaker
260. The audio signal ym (t) emitted from the loudspeaker 260 has the required audio features
and is approaching the ideal sound empirically. Without overcompensating the system, any
unwanted effects of the audio system equipment are removed through the reverse operation of
the precomp filter. As mentioned above, special sound effects can also be introduced into the
resulting audio signal ym (t). The precomp filter system may be implemented as a stand-alone
facility in a digital signal processor or computer, but as mentioned above, has an analog or digital
interface to the subsequent amplifier.
It can also be aimed at incorporating into digital preamplifiers, computer sound cards, compact
stereo systems, home movie systems, computer game consoles or other audio output devices. It is
also possible to implement the precomp filter in a more hardware oriented direction with
customized arithmetic hardware structures. It will be appreciated that the precomp may be
performed separately from sending the audio signal to the actual place of playing the sound. The
precomp signal generated by the precomp filter does not necessarily have to be delivered directly
and immediately from the speech generation system. It may be recorded on another medium for
later sending to the speech generation system. The compensation signal u (t) of FIG. 1 may be
representative of the sound recorded on a CD or DVD disc tailored to the specific audio
equipment and listening environment. It can also be pre-compensated audio files stored on an
Internet server to allow downloading of files from remote locations on the Internet. Finally, the
overall flow of the filter design method according to the exemplary embodiment of the present
invention will be summarized with reference to the flow chart of FIG. This flow chart shows the
pre-steps as well as the actual design steps. This pre-step is preferably used with the present
invention, and therefore an example of the general steps starting with the uncompensated audio
system for designing the precomp filter of the present invention and ending with the
implemented filter and Become. The design starts in step S1. In step S2, a model of the audio
system is determined based on methods known to those skilled in the art. For example, by
determining a model based on physical rules or by performing measurements on an audio
system using known test signals. In step S3, a fixed non-zero filter component is configured. This
configuration may be performed, for example, by using default preset filter components. It can
also be implemented by selecting from a set of pre-set filter components or by inputting fixed
filter components customized by the user's settings. In step S4, weighting is configured. This is
the weighting between approximating the precomp filter to the fixed filter component or
approximating the precompensated model response to the reference system response.
This configuration can be implemented, for example, by using default pre-configured weights, as
in the case of the filter components. It can also be done by choosing from a set of weights or by
entering completely new weights. In step S5, which represents a preferred embodiment of the
present invention, the criterion function comprising the weightings configured in step S4 is
optimized with respect to the adjustable compensator component. This optimization provides a
tunable compensator component that is used to determine the filter parameters of the precomp
filter in step S6 with the fixed non-zero filter component. In step S7, the determined filter
parameters are implemented in the filter hardware or software of the precomp filter. If necessary,
filter parameters must be adjusted. The overall design may be repeated as schematically
represented by dotted line 400 or certain steps may be repeated as represented by dotted line
500. The embodiments described above are examples, and the present invention is not limited
thereto. It is apparent that further modifications, variations and improvements within the spirit of
the present invention disclosed in the present application are included in the scope of the present
invention. <References> [Reference 1] B. ウィドロウ(Widrow)とS.D. Adaptive Signal
Processing, 1985, by Stearns, Prentice-Hall [Ref. 2] S. Adaptive Filter Theory, 3rd Edition,
Prentice-Hall, 1996 by Haykin, Englewood Cliffs, New Jersey [Ref 3] S. T . Neely and J.A. B. Allen,
1979, "Invertibility of a room impulse response", US Pat. H. Hearing Association (J. Acoustical
Society of America), 66, 165-169 [Reference 4] シュテルナド(Sternad)、M.
Johansson and J.A.
Rutstroem's 2000 "Inversion of loudspeaker dynamics by polynomial LQ feedforward control"
(Inversion of loudspeaker dynamics by polynomial LQ feedforward control), IFAC Symposium on
Robust Control Design, Prague, Czech Republic, 2000 6 21st to 23rd of the month [Reference 5]
Sternad and T.W. Soederstroem's 1988 "LQG-optimal feedforward regulators" (LQG-optimal
feedforward regulators), Automatica, Volume 24, pages 557-561 [Ref 6] M. Sternad and A.I.
Ahlen's "LQ control and self-tuning control" in 1993, Polynomial Methods in Optimal Control and
Filtering, Control Engineering Series (Control Engineering Series) ), K. E. Hunt, Chapter 3, Peter
Peregrinus, London [Ref. 7] H. W. Strube's 1980 "Linear prediction on a warped frequency scale",
US Pat. J. Acoustical Society of America, Vol. 68, pages 1071 to 1076 [Reference 8] B. A Course
in H∞ Control Theory, 1987, by A. Francis (Francis), Springer Publishing (Berlin) [Ref. 9] 1985
Control System Synthesis by Vidyasagar, Control System Synthesis, A Factorization Approach,
MIT Press, Cambridge, Mass. [Ref 10] K. J.
Ostrom (Aastroem) and B.B. Computer Control System 3rd Edition 1997 by Wittenmark,
Prentice-Hall, Englewood Cliffs, New Jersey [Ref 11] A. Ahlen and M.W. Sternad's 1991 "Wiener
filter design using polynomials" (Wiener filter design using polynomial equations), IEEE
Transactions on Signal Processing, Vol. 39, pp. 2387-2399 [Reference 12] V . Kucera, 1991
"Analysis and Design of Linear Control Systems" (Analysis and Design of Linear Control Systems),
Academia in Prague and Prentice Hall International, London [Ref 13] W. Bode and C.I. E. 1950
"Simplified derivation of linear least squares smoothing and prediction theory" by Shannon, I. R.
E. Bulletin 38, pages 417 to 425. [Reference 14] A. Ahlen and M.W. Ctern of 1994 in Sternad. T.
レノンデ(Lenondesed. Derivation and design of Wiener filters using polynomials,
Control and Dynamic Systems, Digital Signal Processing and Applications, Academic Press, New
York [Reference 15] B. D. O. Anderson and J.A. B. Moore's 1989 Linear Quadratic Methods,
Optimal Control. Linear Quadratic Methods, Prentice Hall International, London [Ref. 16]
Sternad and A.I. Ahlen, 1993, "A new derivation methodology for polynomial LQ controller
design" (A novel derivation methodology for polynomial LQ controller design), IEEE Transactions
on Automatic Control, Vol. 38, pp. 116-121. The present invention offers the following
advantages: -Tight control of the range and amount of compensation performed by the precomp
filter, providing perfect control over the resulting acoustic response. -Avoid dangerous
overcompensation while providing good compensation that can be achieved without problems.
Demonstrate good compensation capabilities while using a limited number of filter parameters. Get an optimally pre-compensated audio system that produces superior sound quality and
experience results. BRIEF DESCRIPTION OF THE DRAWINGS FIG. 1 is a general illustration of a
system for producing compensated sound. FIG. 2 is a graph (A) showing the amplitude response
of the uncompensated loudspeaker model and the deviation of the phase response of the
uncompensated loudspeaker model as compared to the phase shift of a pure delay It is a graph
figure (B). FIG. 3 shows the discrete time impulse response of the loudspeaker model of FIGS. 2
(A), (B), sampled at 44.1 kHz and delayed by 250 samples for illustration. 4 shows an example of
the impulse response of a scalar FIR compensation filter designed according to the prior art to
reverse the dynamics of the loudspeakers of FIGS. 2 (A), (B) and FIG. 3; 5 shows the impulse
response of a scalar IIR compensation filter designed based on the loudspeaker models of FIGS.
2A, 2B and 3 according to the invention. 6A is a graph showing the amplitude response of the
loudspeaker model of FIG. 2A compensated by the IIR filter of FIG. 5A and the loudspeaker model
of FIG. 2B compensated by the IIR filter of FIG. 5; (B) is a graph showing the deviation of the
phase response of the signal in comparison with the phase shift of a pure delay. 7 shows the
compensated impulse response of the loudspeaker model of FIG. 3 compensated by the IIR filter
of FIG. 5; 8 shows the frequency response amplitude of the weighting function used in the design
of the IIR filter of FIG. 5;
9 shows the compensated impulse response of FIG. 8 when using compensation without control
penalty. FIG. 10 is a graph (A) showing the amplitude response of the prior art FIR filter
compensated loudspeaker model of FIG. 2 of FIG. 4 and FIG. 2 (compacted with the prior art FIR
filter of FIG. FIG. 7B is a graph (B) showing the deviation of the phase response of the
loudspeaker model of B) compared to the phase shift of a pure delay. FIG. 11 is a schematic
diagram illustrating a specific embodiment of a filter design structure according to the present
invention. FIG. 12 is a block diagram of a computer based system that may be used to practice
the present invention. FIG. 13 is a diagram showing an audio system incorporating a precomp
filter set by the design method of the present invention. FIG. 14 is a process flow diagram
showing an overall flow of a filter design method according to an exemplary embodiment of the
present invention.
Без категории
Размер файла
55 Кб
Пожаловаться на содержимое документа