**ALIASING REDUCTION USING COMPLEX-EXPONENTIAL MODULATED FILTERBANKS**

Ekstrand, Per (Södermannagatan 45 Stockholm, S-116 40, SE)

*;*

**G10L11/00***;*

**H03H17/02***;*

**H03H21/00***; (IPC1-7): H03H17/00*

**H03M7/30**WO2001041320A1 |

US5732189A | ||||

US5285498A | ||||

US5568142A |

1. | A method for estimation of energy measures of coefficients or subband signals obtained from a digital filterbank, characterized by: optimizing a symmetric lowpass prototype filter po (n), having filter order N, not constrained to have the perfect reconstruction property; building an Mchannel analysis filterbank by complexexponential modulation of said prototype filter; where said filterbank has filter coefficients as and n = 0, 1,..., N, and k= 0, 1,..., M1 ; filtering a realvalued time domain signal through said filterbank ; and calculating the squared absolute values of the complexvalued subband signals obtained from said filtering. |

2. | A method according to claim 1, characterized in that said analysis filterbank is used to estimate average energy measures as where vk (n) is said complexvalued subband signal of channel k, and w (n) is a window of length 2L1 centered around n = 0. |

3. | A method for reduction of aliasing emerging from modifications of the coefficients or subband signals obtained from a digital filterbank, characterized by: optimizing a symmetric lowpass prototype filter po (n), having filter order N, not constrained to have the perfect reconstruction property; building an Mchannel filterbank by complexexponential modulation of said prototype filter; where said filterbank has analysis and synthesis filter coefficients as andn=0... N, andk=0... M1 ; filtering a realvalued time domain signal through the analysis part of said filterbank; modifying the complexvalued subband signals obtained from said filtering; filtering said modified complexvalued subband signals through the synthesis part of said filterbank; and taking the real part of the complexvalued time domain output signal, where said output signal is the sum of the signals obtained from said synthesis filtering. |

4. | A method according to claims 13, characterized in that said lowpass prototype filter has a filter order N that is higher than 2M1, where Mis the number of channels in said digital filterbank. |

5. | A method according to claims 14, characterized in that said optimization of said lowpass prototype filter is accomplished by minimizing a composite objective function £ tot (a) as #tot(α)=α#t+(1α)#a where a is a weighting constant, c t is the passband error and s a is the aliasing error. |

6. | A method according to claim 5, characterized in that said passband error £ t iS calculated as and said aliasing error £ a iS calculated as where Al (ej#), l= 0... Mol2, are the alias gain terms evaluated on the unit circle. |

7. | A method according to claim 6, characterized in that said alias gain terms are calculated using gain adjusted versions of said analysis and synthesis filters according to hk(eq)(n)=gkhk(n) and fk(eq)(n)=gk fk(n), where gk represents randomized or deterministic realvalued gainfactors, n = 0... N, and k = 0... M1. |

8. | A method according to claims 57, characterized in that said lowpass prototype filter has a filter order of at least a number N, where Nis chosen so that said composite objective function is smaller than a predefined value. |

9. | A method according to claims 12, characterized in that said analysis filterbank is used for the estimation of energy measures in a high frequency reconstruction system. |

10. | A method according to claims 38, characterized in that said complexexponential modulated filterbank is used as the envelope adjusting filterbank in a high frequency reconstruction system. |

11. | A method according to claims 38, characterized in that said complexexponential modulated filterbank is used as the filters in a digital equalizing system. |

12. | A method according to claims 38, characterized in that said complexexponential modulated filterbank is used in an adaptive filtering system. |

13. | A method according to claims 38, characterized in that said complexexponential modulated filterbank is used in a multiband compander system. |

14. | An apparatus for estimation of energy measures of coefficients or subband signals obtained from a digital filterbank, characterized by: means for optimizing a symmetric lowpass prototype filter po (n), having filter order N, not constrained to have the perfect reconstruction property ; means for building an Mchannel analysis filterbank by complexexponential modulation of said prototype filter; where said filterbank has filter coefficients as and n = 0, 1,..., N, and k = 0, 1,..., M1 ; means for filtering a realvalued time domain signal through said filterbank; and means for calculation of the squared absolute values of the complexvalued subband signals obtained from said filtering. |

15. | An apparatus for reduction of aliasing emerging from modifications of the coefficients or subband signals obtained from a digital filterbank, characterized by: means for optimizing a symmetric lowpass prototype filter po (n), having filter order N, not constrained to have the perfect reconstruction property; means for building an Mchannel filterbank by complexexponential modulation of said prototype filter; where said filterbank has analysis and synthesis filter coefficients as and n = 0, 1,..., N, and k = 0, 1,..., M1 ; means for filtering a realvalued time domain signal through the analysis part of said filterbank; means for modifying the complexvalued subband signals obtained from said filtering; means for filtering said modified complexvalued subband signals through the synthesis part of said filterbank ; and means for taking the real part of the complexvalued time domain output signal, where said output signal is the sum of the signals obtained from said synthesis filtering. |

Ferreira, J. M. N. Viera, AES preprint, 98th Convention 1995 February 25-28 Paris, N. Y., USA], adaptive filters [Adaptive Filtering in Subbands with Critical Sampling: Analysis, Experiments, and Application to Acoustic Echo Cancellation"A. Gilloire, M. Vetterli, IEEE Transactions on Signal Processing, vol. 40, no. 8, August, 1992], multiband companders, and to audio coding systems using high frequency reconstruction (HFR), where a digital filterbank is used for the adaptive adjustment of the spectral envelope, such as the Spectral Band Replication (SBR) system [WO 98/57436].

BACKGROUND OF THE INVENTION A digital filter bank is a collection of two or more parallel digital filters. The analysis filter bank splits the incoming signal into a number of separate signals named subband signals (or spectral coefficients). The filter bank is critically sampled (or maximally decimated) when the total number of subband samples per unit time is the same as that for the input signal. The synthesis filter bank combines these subband signals into an output signal. A popular type of critically sampled filterbanks is the cosine modulated filterbank.

The filters in the cosine modulated system are obtained by cosine modulation of a low-pass filter, a so- called prototype filter. The cosine modulated banks offer very effective implementations and are often used in natural audio codecs ["Introduction to Perceptual Coding"K. Brandenburg, AES, Collected Papers on Digital Audio Bitrate Reduction, 1996]. However, any attempt to alter the subband samples or spectral coefficients, e. g. by applying an equalizing gain curve or quantizing the samples, renders severe aliasing artifacts in the output signal.

SUMMARY OF THE INVENTION The present invention shows that impairments emerging from modifications of the subband signals can be significantly reduced by extending a cosine modulated filterbank with an imaginary sine modulated part, forming a complex-exponential modulated filterbank. The sine extension eliminates the main alias terms present in the cosine modulated filterbank. Moreover, the invention presents a method, referred to as alias term minimization (ATM), for optimization of the prototype filter. The complex-exponential modulation creates complex-valued subband signals that can be interpreted as the analytic signals of the signals obtained from the real part of the filterbank, i. e. the underlying cosine modulated filterbank. This feature provides an inherent measure of the instantaneous energy of the subband signals.

The main steps for operation of the complex-exponential modulated filterbank according to the present invention are: 1. The design of a symmetric low-pass prototype filter with cutoff frequency 7r/2M, optimized for a desired aliasing rejection and passband flatness; 2. The construction of an M-channel filterbank by complex-exponential modulation of the optimized prototype filter ; 3. The filtering of a real-valued time domain signal through the analysis part of the filterbank ; 4. The modification of the complex-valued subband signals according to a desired, possibly time- varying, equalizer setting; 5. The filtering of the modified complex-valued subband samples through the synthesis part of the filterbank ; and 6. The computation of the real part of the complex-valued time domain output signal obtained from the synthesis part of the filterbank.

The most attractive applications of the invention are improvement of various types of digital equalizers, adaptive filters, multiband companders and adaptive envelope adjusting filterbanks used in HFR systems.

BRIEF DESCRIPTION OF THE DRAWINGS The present invention will now be described by way of illustrative examples, not limiting the scope nor the spirit of the invention, with reference to the accompanying drawings, in which: Fig. l illustrates the analysis and synthesis sections of a digital filterbank ; Fig. 2 shows the magnitudes in a composite alias component matrix of a cosine modulated filterbank ; Fig. 3 shows the magnitudes in a composite alias component matrix of a complex-exponential modulated filterbank ; Fig. 4 illustrates wanted terms and main alias terms in a cosine modulated filterbank adjusted for a band- pass filter response ; Fig. 5 shows the attenuation of alias gain terms for different realizations of complex-exponential modulated filterbanks ; Fig. 6 illustrates the analysis part of a complex-exponential modulated filterbank system according to the present invention; and Fig. 7 illustrates the synthesis part of a complex-exponential modulated filterbank system according to the present invention.

DESCRIPTION OF PREFERRED EMBODIMENTS It should be understood that the present invention is applicable on a range of implementations that incorporates digital filterbanks other than those explicitly mentioned in this patent.

Digital Filterbanks A digital filter bank is a collection of two or more parallel digital filters that share a common input or a common output ["Multirate Systems and Filter Banks"P. P. Vaidyanathan Prentice Hall: Englewood Cliffs, NJ, 1993]. When sharing a common input the filter bank is called an analysis bank. The analysis bank splits the incoming signal into M separate signals called subband signals. The analysis filters are denoted Hk (z), where k = 0... M-1. The filter bank is critically sampled (or maximally decimated) when the subband signals are decimated by a factor M. The total number of subband samples per unit time is then the same as the number of samples per unit time for the input signal. The synthesis bank combines these subband signals into a common output signal. The synthesis filters are denoted Fk (z), for k = O... M-1. A maximally decimated filter bank with M channels (subbands) is shown in Fig. 1. The analysis part 101 produces the signals Vk (z), which constitute the signals to be transmitted, stored or modified, from the input signal X (z). The synthesis part 102 recombines the signals Vk (z) to the output signal X (z).

The recombination of Vk (z) to obtain the approximation X (z) of the original signal X (z) is subject to several errors. One of these is aliasing, due to the decimation and interpolation of the subbands. Other errors are phase and amplitude distortion.

Following the notations of Fig. l, the outputs of the analysis filters Hk (z) 103 are Xk (z)=Hk(z)X(z) (1) where k = 0... M-1. The decimators 104 give the outputs where W = e-i 21rIM. The outputs of the interpolators 105 are given by and the sum of the signals obtained from the synthesis filters 106 can be written as

is the gain for the/ alias term X (zW). Eq. (4) can be rewritten as The last sum on the right hand side (RHS) constitutes the sum of all non-wanted alias terms. Canceling all aliasing, that is forcing this sum to zero by means of proper choices of Ht (z) and Fk (z), gives is the overall transfer function or distortion function. Choosing the synthesis filter Fk (z) so that Fk (z) = z NHk (Z) (9) where N is the order of the analysis filter, results in the following transfer function: The notation H (z) is the Z-transform of the time-reversed and complex-conjugated sequence hk (n).

Eq. (10) evaluated on the unit circle gives Eq. (l 1) shows that T (z) has linear phase, and thus has no phase distortion. Further, if the last sum on the RHS is a constant, there is no amplitude distortion. The overall transfer function is in this case simply a delay with a constant scale factor c, i. e.

T (z) = cz-N (12) which substituted into Eq. (7) gives #(z)=cz-NX(z) The type of filters that satisfy Eq. (13) are said to have the perfect reconstruction (PR) property.

Cosine Modulated Filterbanks In a cosine modulated filterbank the analysis filters hk (n) are cosine modulated versions of a symmetric low-pass prototype filter po (n) as

where M is the number of channels, k = 0... M-1, Nis the prototype filter order and n = 0... N. The sum of the real-valued prototype filter coefficients is assumed to be unity as Following the same notation, the synthesis filters are given by The analysis filter bank produces real-valued subband samples for real-valued input signals. The subband samples are downsampled by a factor M, which makes the system critically sampled. Depending on the choice of the prototype filter, the filterbank may constitute a near perfect reconstruction system, a so called pseudo QMF bank [US5436940], or a perfect reconstruction (PR) system. An example of a PR system is the modulated lapped transform (MLT) ["Lapped Transforms for Efficient Transform/Subband Coding"H. S. Malvar, IEEE Trans ASSP, vol. 38, no. 6,1990]. One inherent property of the cosine modulation is that every filter has two passbands; one in the positive frequency range and one corresponding passband in the negative frequency range.

Eq. (5) can be written in matrix form as a=Hf (17) or explicitly as The matrix H is called the alias component (AC) matrix. For a closer examination of this equation, f may be written as or in a compact form as

f = F e (20) Substituting Eq. (20) in Eq. (17), the alias gains can be written as a = HF e, where the product HF = U (21) is a M x M matrix, hereinafter referred to as the composite alias component matrix.

For a cosine modulated system, the dominant terms in the composite alias component matrix are the first row and four diagonals. The three-dimensional plot of Fig. 2 illustrates the magnitudes of the components in this matrix. The first row holds the terms from the transfer function, Eq. (8), while the four diagonals primarily consist of the main alias terms, i. e. the aliasing due to overlap between filters and their closest neighbors. It is easily seen that the main alias terms emerge from overlap in frequency between either the filters negative passbands with frequency modulated versions of the positive passbands, or reciprocally, the filters positive passbands with frequency modulated versions of the negative passbands. Summing the terms of the rows in the composite alias component matrix, i. e. calculating the alias gains, results in cancellation of the main alias terms. The aliasing is canceled in a pairwise manner, where the first main alias term is canceled by the second in the same row. Superimposed on the main alias terms are also other smaller alias terms. If the prototype filter characteristics is so that the transition-and stop-band of the filters have substantial overlap with their modulated versions, these alias terms will be large. As an example, the second and the last row consists of alias terms induced by the overlap of filters with their closest modulated versions. For a PR system, these smaller alias terms also cancels completely when summing the terms for the alias gains. In the pseudo QMF system, however, these terms remain.

Complex-Exponential Modulated Filterbanks Extending the cosine modulation to complex-exponential modulation according to the present invention yields the analysis filters hk (n) as using the same notation as before. This can be viewed as adding an imaginary part to the real-valued filterbank, where the imaginary part consists of sine modulated versions of the same prototype filter.

Considering a real-valued input signal, the output from the filter bank can be interpreted as a set of subband signals, where the real and the imaginary parts are Hilbert transforms of each other. The resulting subbands are thus the analytic signals of the real-valued output obtained from the cosine modulated filterbank. Hence, due to the complex-valued representation, the subband signals are oversampled by a factor two.

The synthesis filters are extended in the same way as Eq. (22) and (23) implies that the output from the synthesis bank is complex-valued. Using matrix notation, where Ca is a matrix with analysis filters from Eq. (14), and Sa is a matrix with filters as

the filters of Eq. (22) is obtained as Ca + j Sa. In these matrices, k is the row index and n is the column index. Analogously, the matrix Cs has synthesis filters from Eq. (16), and Ss is a matrix with filters as Eq. (23) can thus be written Cs +j Ss, where k is the column index and n is the row index. Denoting the input signal x, the output signal y is found from y=(Cs+j Ss)(Ca+j Sa) x = (CsCa - SsSa) x +j (CsSa + SsCa) x (26) As seen from Eq. (26), the real part consists of two terms; the output from the ordinary cosine modulated filterbank and an output from a sine modulated filterbank. It is easily verified that if a cosine modulated filterbank has the PR property, then its sine modulated version, with a change of sign, constitutes a PR system as well. Thus, by taking the real part of the output, the complex-exponential modulated system offers the same reconstruction accuracy as the corresponding cosine modulated version.

The complex-exponential modulated system can be extended to handle also complex-valued input signals. By extending the number of channels to 2M, i. e. adding the filters for the negative frequencies, and keeping the imaginary part of the output signal, a pseudo QMF or a PR system for complex-valued signals is obtained.

Examining the composite alias component matrix from Eq. (21), the main alias diagonals vanish for the complex-exponential modulated filterbank. This is easily understood since the complex-exponential modulated filterbank has only one passband for every filter. In other words, the filterbank is free from main alias terms, and do not rely on the pairwise aliasing cancellation mentioned above. The composite alias component matrix has the dominant terms on the first row only. Fig. 3 shows the magnitude of the components in the resulting matrix. Depending on the prototype filter characteristics, the terms on rows 1 through M-1, are more or less attenuated. The absence of main alias terms makes the aliasing cancellation constraint from the cosine (or sine) modulated filterbank obsolete in the complex-exponential modulated version. Both the analysis and synthesis filters can thus be found from since po (n) = po (N-n) for a symmetric prototype filter. As before, Mis the number of channels, k= 0... M-1, Nis the prototype filter order and n = 0... N.

Referring to Eq.(4), the Z-transform of the real part of the output signal #(n) is <BR> <BR> <BR> Z{Re(#(n))}= XR(z)=X(z)+X*(z) (28)<BR> 2

The notation * (z) is the Z-transform of the complex-conjugated sequence x (n). From Eq. (4), it follows that the transform of the real part of the output signal is where it is used that the input signal x (n) is real-valued. Eq. (29) may after manipulation be written 2MXR (z) = X (z) {Ao (z) + A0* (Z)} + +X (-z) {AMx2 (Z) + AM/2* (Z)} + By inspecting Eq. (30), and recalling the transform of Eq. (28), it is obvious that the real part of ao (n) must be a dirac pulse for a PR system. Moreover, the real part of aM/2 (n) must be zero and the alias gains, for l= 1... M/2-1, must satisfy AM-l (z) =-) (31) In pseudo QMF systems, Eq. (31) holds true only approximately. Moreover, the real part of ao (n) is not exactly a dirac-pulse, nor is the real part of aM/2 (n) exactly zero.

Modification of subband signals Changing the gains of the channels in a cosine modulated filterbank, i. e. using the analysis/synthesis system as an equalizer, renders severe distortion due to the main alias terms. Say for example that the intention is to adjust an eight-channel filterbank for a band-pass response, where except for the second and third channel all subband signals are set to zero. The composite alias component matrix from Eq. (21) is then an 8 x 8 matrix where all elements are zero except for the elements on the second and third column, Fig. 4. There are seven significant alias terms left as indicated in the figure. The aliasing from rows three and five will be canceled, since the main alias terms in these rows have the same gains, i. e. the pairwise cancellation is working intentionally. In rows two, four and six however, there are single alias terms, since their corresponding alias terms have zero gain. The alias cancellation will hence not work as intended, and the aliasing in the output signal will be substantial.

From this example it is obvious that a substantial improvement is achieved when using complex- exponential modulated filterbanks as equalizers. The 8-channel system depicted in Fig. 4 has a prototype filter of order 128. The total aliasing attenuation is only 16 dB in the above equalizer example. Moving to complex-exponential modulation gives an aliasing attenuation of 95 dB. Due to the non-existing main

alias terms, the resulting aliasing is dependent only on the suppression of the alias terms emanating from overlap between filters and their modulated versions. It is thus of great importance to design the prototype filter for maximum suppression of the alias gains terms. The first term on the RHS of Eq. (30) evaluated on the unit circle gives the error energy et of the transfer function as The energy of the total aliasing ea can be calculated by evaluating all remaining terms on the RHS of Eq. (30) on the unit circle as Due to symmetry, Eq. (9) and the fact PO(z)=z-N#0(z) the terms within the curly braces of the sum in Eq. (33) are equal. The total aliasing energy thus has M/2-1 terms as The minimization of the alias gain terms is done by optimizing the prototype filter. This is preferably accomplished by minimizing a composite objective function, using standard nonlinear optimization algorithms, for example the Downhill Simplex Method ["Numerical Recipes in C, The Art of Scientific Computing, Second Edition"W. H. Press, S. A. Teukolsky, W. T. Vetterling, B. P. Flannery, Cambridge University Press, NY, 1992]. For alias term minimization (ATM) of the prototype filter according to the invention, the objective function looks like ctot (α)=α#t+(1-α)#a (36) During the optimization, a random equalization curve is applied to the filterbank when calculating fia, i. e. the analysis and synthesis filters are multiplied with gainfactors gk as

Hkeq) (z) =gkHk (z) and Fk (eq) (z) =gkFk (z) (37) and the resulting filters Hk (e) and Fk (eq), k = 0... M-1, are used when calculating the alias gain terms Al (z) forl=1... M1.

In Fig. 5, the alias gains of five different complex-exponential modulated systems are compared. Four of these are 8-channel systems and one is a 64-channel system. All of the systems have prototype filter lengths of 128. The dotted trace and the solid trace with stars shows alias components for two pseudo QMF systems, where one is alias term minimized. The dashed and the dash-dotted traces are the components for two 8-channel perfect reconstruction systems, where again one of the systems is alias term minimized. The solid trace is the alias components for a complex-exponential modulated lapped transform (MLT). All the systems are adjusted for band-pass responses according to the example above, and the results are tabulated in Table 1. The rejection of total aliasing is calculated as the inverse of Eq. (33). The passband flatness is calculated as the inverse of Eq. (32) with the integration interval adjusted for the bandpass response.

Table 1 System Rejection of total aliasing Passband flatness 8-ch Pseudo QMF ATM N=128 114,7 dB 98, 1 dB 8-ch Pseudo QMF N=128 95,4 dB 87, 6 dB 8-ch PR ATM N=127 77, 3 dB 132, 7 dB 8-ch PR N=127 55, 0 dB 93, 6 dB 64-ch MLT N=127 38, 5 dB 87, 1 dB As seen from the numbers in Table 1, a substantial improvement is achieved when moving from the 64- channel MLT to the 8-channel PR systems. The MLT is a perfect reconstruction system and has only (N+1) l 2M= 1 coefficient per polyphase component. The number of coefficients for the 8-channel PR systems are 128/16 = 8. This enables filters with higher stopband attenuation and higher rejection of alias terms. Moreover, it is seen that alias term minimization of the PR system rejects the aliasing and improves the passband flatness significantly. Comparing the pseudo QMF systems and the PR systems, it is clear that the aliasing rejection improves by 40 dB while almost preserving the passband flatness. An additional alias rejection of approximately 20 dB and improved passband flatness of 10 dB is achieved when minimizing the alias terms. Thus, it is clear that the perfect reconstruction constraint imposes limitations to a filterbank used in an equalization system. A pseudo QMF system can always be designed for adequate reconstruction accuracy, since all practical digital implementations have limited resolution in the numeric representation. For both the pseudo QMF and the PR system, it is obvious that an optimum system is built on a prototype filter that has large rejection of the stopband. This imposes the usage of prototype filters with relative lengths longer than the windows used in the MLT.

A great advantage of the complex-exponential modulated system is that the instantaneous energy is easily calculated since the subband signals constitute the analytic signals of the real-valued subband signals

obtained from a cosine modulated filterbank. This is a property of great value in for example adaptive filters, automatic gain controls (AGC), in multiband companders, and in spectral band replication systems (SBR), where a filterbank is used for the spectral envelope adjustment. The averaged energy within a subband k might be calculated as where vk (n) is the subband samples of channel k, and w (n) is a window of length 2L-1 centered around n = 0. This measure is then used as an input parameter for the adaptation or gain-calculation algorithms.

Practical implementations Using a standard PC or DSP, real-time operation of a complex-exponential modulated filterbank is possible. The filterbank may also be hard-coded on a custom chip. Fig. 6 shows the structure for an effective implementation of the analysis part of a complex-exponential modulated filterbank system. The analogue input signal is first fed to an A/D converter 601. The digital time domain signal is fed to a shift register holding 2M samples shifting M samples at a time 602. The signals from the shift register are then filtered through the polyphase coefficients of the prototype filter 603. The filtered signals are subsequently combined 604 and in parallel transformed with a DCT-IV 605 and a DST-IV 606 transform.

The outputs from the cosine and sine transforms constitute the real and the imaginary parts of the subband samples respectively. The gains of the subband samples are modified according to the current spectral envelope adjuster setting 607.

An effective implementation of the synthesis part of a complex-exponential modulated system is shown in Fig. 7. The subband samples are first multiplied with complex-valued twiddlefactors 701, and the real part is modulated with a DCT-IV 702 and the imaginary part with a DST-IV 703 transform. The outputs from the transforms are combined 704 and fed through the polyphase components of the prototype filter 705. The time domain output signal is obtained from the shift register 706. Finally, the digital output signal is converted back to an analogue waveform 707.

The above-described embodiments are merely illustrative for the principles of the complex-exponential modulated filterbank systems according to the present invention. It is understood that modifications and variations of the arrangements and the details described herein will be apparent to others skilled in the art.

It is the intent, therefore, to be limited only by the scope of the impending patent claims and not by the specific details presented by way of description and explanation of the embodiments herein.

**Previous Patent:**SURFACE ACOUSTIC WAVE COMPONENT

**Next Patent: TWO-DIMENSIONAL PYRAMID FILTER ARCHITECTURE**