This application claims priority of International Application No. PCT/EP2004/008217, filed Jul. 23, 2004 and German Application No. 103 34 357.1, filed Jul. 25, 2003, the complete disclosures of which are hereby incorporated by reference.
a) Field of the Invention
The present invention is directed to a method and apparatus for the digitization of analog signals and a method and apparatus for data compression of analog signals.
b) Description of the Related Art
In many fields of digital information transmission there exists the common goal of transporting an initially analog source signal (measuring, studio-quality audio signal processing, and so on) to a receiver while making use of the advantages of digital transmission in order to reproduce this analog source signal in the receiver in the form of an analog output signal. For this purpose, the analog signal must be digitized and quantized on the one hand and a corresponding data compression must be carried out on the other hand in order to transmit the signal with the highest possible quality via a digital transmission system with a limited rate. After a PCM encoding, methods for data compression are usually applied which make use of the redundancy within the source signal and irrelevance with respect to specific characteristics of the consumer of the signal. The numerous existing methods for digitizing analog source signals use approaches that sometimes differ fundamentally, but all of them can be divided into two main categories:
i) Methods in which the reconstructed waveform very closely approximates the original waveform, i.e., there is no use of irrelevance. The term “waveform-preserving” is used here in place of the often applied term “lossless” with respect to the waveform coding because the digitization of an analog signal at a limited data rate is, on principle, not “lossless” (i.e., infinite entropy of a continuous random variable).
ii) Non-waveform-preserving methods. These methods are of crucial importance, for example, in the audio field, where often only the subjective aural impression at the receiver output is decisive (e.g., use of psychoacoustic masking effects). In this case, waveform changes (amplitude distortion and phase distortion) undergone by the signal through quantization and compression generally only play a subordinate role. Normally, a signal processing using irrelevance leads to a reconstructed waveform that differs greatly from the original waveform and, moreover, is often afflicted by extensive signal delay (e.g., because of spectral transformations or equivalent block-based methods). The transmission quality in a method of the type mentioned above cannot be measured by a 10 log10 signal/noise ratio (SNR) in the conventional way; rather, it must be determined by time-consuming listening tests by trained personnel. However, signal coding methods of this type are totally unusable for many areas of application (e.g., metrology, recording of waveforms for further signal processing at a later time, real-time signal transmission using digital modulation methods that do not permit significant signal delay, e.g., for wireless digital stage microphones).
Therefore, it is the primary object of the present invention to provide a method and apparatus for the digitization and compression of analog signals with improved quality.
This object is met by a method for the digitization of analog signals comprising the steps of digitizing analog source signals, transforming the digitized source signals from the time domain to the spherical domain, wherein the transformation is a D-dimensional transformation with D>2 and logarithmic quantizing of the radius in the spherical domain. It is also met by an apparatus for the digitization of analog signals comprising means for the digitization of analog source signals, means for the transformation of the digitized source signals from the time domain to the spherical domain, wherein the transformation is a D-dimensional transformation with D>2, and means for the logarithmic quantization of the radius in the spherical domain. It is further met by a method for the compression of analog signals comprising the steps of digitizing analog source signals by the method described above and carrying out a differential pulse code modulation. It is also met by an apparatus for the compression of analog signals comprising the apparatus described above including an encoder for differential pulse code modulation and a forward prediction device for determining a starting value for the samples of the quantization based on the current state of a prediction filter, a reconstruction device for reconstructing the subsequent D samples, a D-dimensional logarithmic spherical quantization device for quantizing the values obtained by the forward prediction in order to determine a starting cell, wherein the prediction of the differential pulse code modulation is run through iteratively in order to determine a quantization cell with the smallest quantization error.
Accordingly, a method is provided for the digitization of analog source signals in which a D-dimensional spherical logarithmic quantization of the analog source signals is carried out.
According to one aspect of the present invention, a method and an apparatus are likewise provided for the compression of analog source signals in which a digitization of analog source signals is carried out in accordance with the method described above, and wherein a differential pulse code modulation is carried out.
Accordingly, a waveform-preserving method is provided which meets the following requirements: a) low data rate through the use of favorable packing characteristics of multi-dimensional lattices (vector quantization) and through the use of dependencies in the sequence of samples from the source signal; b) extremely high dynamic range, i.e., the SNR is constant over a very large modulation range of, e.g., 60 dB or more with respect to the short-term variance of an analog source signal; c) high objectively measurable signal/noise ratio by segments for short segments of samples, preferably within the meaning of the ratio of variance of the useful signal to the mean square error; d) insensitivity to specific signal parameters such as probability density function by segment, etc.; and, most importantly, e) introduction of an extremely small signal delay on the order of a few (up to ten) sampling periods.
The invention is based on the idea of providing a method for high-resolution, waveform-preserving digitization of analog signals, wherein conventional scalar logarithmic quantization is transferred to multi-dimensional spherical coordinates, and the advantages resulting from this, e.g., a constant signal/noise ratio over an extremely high dynamic range with very low loss with respect to the rate-distortion theory. In order to make use of the statistical dependencies present in the source signal for an additional gain in the signal/noise ratio, the differential pulse code modulation (DPCM) is combined with spherical logarithmic quantization. The resulting method achieves an effective data reduction with a high long-term and short-term signal/noise ratio with an extremely small signal delay.
Further embodiments of the invention are indicated in the subclaims.
In the following, the invention will be described more fully with reference to the accompanying drawings.
In the drawings:
The aim of the logarithmic quantization is a high dynamic of the quantizer, that is, a wide range of the average signal level in which the SNR and, therefore, the maximum relative quantization error
is constant, wherein Δqi is the width of the ith quantization interval and ri is the respective reconstruction value. This aim leads to logarithmic quantization, e.g., within the meaning of the A-law according to K. Tröndle, R. Weiβ, Introduction to Pulse Code Modulation, Oldenbourg Verlag, Munich, 1974. For the range of average modulation, the following signal/noise ratio is given for R>>1 according to N. S. Jayant, P. Noll, Digital Coding of Waveforms, Prentice-Hall, Englewood Cliffs, N.J., 1984:
where R is the average rate of quantization (bits/sample) and A represents the usual parameters of logarithmic quantization; that is, the maximum relative error is constant for samples with an amount greater than 1/A (with reference to a signal range of the quantization of −1 to +1). Therefore, a signal/noise ratio according to (1) is achieved after an average signal level of the source signal of
−B1:=20 log10(1/A) dB. (2)
Accordingly, the value B1 characterizes the dynamic range of the logarithmic quantization. Also, the signal/noise ratio is entirely independent from the probability density function (PDF) of the source signal within this dynamic range so that the method can be used universally. However, in comparison with a uniform quantization at full drive by an equidistributed signal, the term 10 log10(3/(1+ln(A))2)to a certain extent represents a signal/noise ratio loss by companding, which is the price of a large dynamic range.
A quantization in spherical coordinates in D dimensions will be described in the following.
Spherical logarithmic quantization belongs to the family of vector quantization methods (e.g., R. M. Gray, David L. Neuhoff, Quantization, IEEE Transactions on Information Theory, pp. 2325-2383, October 1998, Manfred Herbert, Lattice Quantization of Speech Signals and Speech Model Signals. Selected Works on Information Systems, No. 79, editor: H.-W. Schüβler, Erlangen 1991). By means of dense spherical packing in many dimensions, considerable gains in signal/noise ratios can be achieved in the quantization even when no statistical dependencies within the signals can be made use of. In spherical quantization, a vector x:=(x1, . . . , xD) with D samples in Cartesian coordinates is expressed in polar coordinates u:=(φ1, . . . , φD−1,r). The D−1 angle φ1 and the radius r are given by the following equations (where j is an imaginary unit and arg (·) is an argument function which indicates the angle of a complex number in radian measure):
The reconstruction of the Cartesian components from vector u is carried out from this according to
xi=r·bi−1·sin(φi−1), i ε {D, . . . , 2} (6)
x1=r·b1·cos(φ1)=r·b0, (7)
with the radii bi of the “parallel of latitude” of the unit sphere (radius 1):
Accordingly, an economical recursive representation is found proceeding from φD−1. An efficient implementation of forward transformation and backward transformation is made possible with little complexity by means of the so-called CORDIC algorithm (described in Volder, Jack, The CORDIC Trigonometric Technique, IRE Trans. Electronic Computing, Vol. EC-8, pp. 330-334, September 1959).
Spherical logarithmic quantization is described in the following.
In order to continue to make use of the advantages of logarithmic quantization, i.e., the independence of the SNR from the PDF of the source signal, with quantization in polar coordinates, the amount (radius) is quantized logarithmically according to the rules of the A-law. The angles φi are uniformly quantized individually, in each instance with the quantization intervals being selected in each instance depending on the higher-order angles {circumflex over (φ)}l, lε{i+1,l+2, . . . , D−1} that have already been quantized and which are also used for the reconstruction. In this way, a very simple implementation of the quantization and the signal reconstruction is achieved and a stepwise processing of the individual coordinates, as in scalar quantization, is preserved by the iterative procedure according to (3) to (9).
Due to the proportionality of the length of an arc segment to the radius, the requirements for the logarithmic quantization, i.e., the proportionality of the size of a quantization interval to the signal value, are already met for uniform quantization of the angles.
In order to facilitate the implementation of quantization in spherical coordinates, we start from approximately cubic quantization cells, i.e., the surface of a D-dimensional sphere with radius 1 is quantized by a network of D−1 dimensional (hyper-)cubes. As will be shown in the following, however, this suboptimal quantization of the spherical surface compared to an optimally densely packed lattice results in a loss of only
in the limit D→∞, that is, a rate loss of about 1/4 bit/sample, which is accepted for the sake of facilitated implementation.
The following description explains how the MD (i.e., M:=2R) quantization stages available per quantization step are optimally divided into the individual quantization intervals for the radius and the surface of the D-dimensional unit (hyper-)sphere.
As is conventional, the applicable compressor characteristic for the logarithmic range
of the A-law rule is
This is derived by
A refers to the free parameter of the A-law and r0 is introduced as a normalization factor for the radius which will be described more fully in the following. With MD quantization intervals for the radius (Dth component of vector u), the width of the quantization cell in direction of the radius is given by:
It will be noted that Δr(r) is not dependent on r0 in the domain in question.
In order that a saturation behavior similar to that in Cartesian coordinates can be obtained with quantization in polar coordinates, the maximum value r0 for the radius is normalized in such a way that the D-dimensional sphere has the same volume as a D-dimensional cube with edge length 2 (corresponding to a quantization range xiε[−1;1] in every dimension), i.e.,
with the volume αD of the D-dimensional unit sphere according to J. H. Conway and N. J. A. Sloane, Sphere Packings, Lattices, and Groups, Springer-Verlag, 3rd edition.
To facilitate the determination of the resulting signal/noise ratio, r=1 is used in the following, i.e., those quantization cells lying on the surface of a sphere with unit radius are considered. This is possible without limiting generality because the SNR is not dependent upon radius due to the logarithmic quantization in the domain:
The width of the quantization cells is now given by:
By means of the uniform angular quantization mentioned above, the surface of the D-dimensional unit sphere is divided into Mφ cells having in each instance the form of D−1-dimensional cubes. For the surface of this D−1-dimensional sphere, according to J. H. Conway and N. J. A. Sloane, Sphere Packings, Lattices, and Groups, Springer-Verlag, 3rd edition:
S=βD·rD−1, where βD=D·αD. (16)
The division of the quantizing stages into radius and angle is described the following.
The following considerations must be taken into account for a fair division of the MD available quantization cells into radius and surface of the unit sphere: with respect to the approximation by cubic quantization cells, i.e., approximately equal dimensioning of the quantization cells in all dimensions (radius and arc segment), the surface of the unit sphere (r=1) must be divided into Mφ equal (D−1)-dimensional cubes with identical edge length Δ.
Since there are M=2R available intervals per sample,
must be demanded in addition.
With (15), (17) and (18), the quantity of the quantization intervals available for the quantization of the radius is given by:
The quotients MD/(M/2) of the quantity of quantization cells dedicated to the radius are plotted in
Consequently, the resolution is increased by Δn bits for small values of r.
The edge length of the quantization cells on the surface of the unit sphere is obtained using (15):
The constant c is determined by the selected dynamic range, see (2) and (10).
The actual uniform quantization of the angle variables for identical arc segments on the unit sphere is obtained by means of the recursive equations (8) and (9). Mi intervals are available for the coordinate φi:
where └x┘εIN is the greatest whole number≦x with xεIR+ and {circumflex over (b)}i corresponding to (8) and (9) for quantized angles {circumflex over (φ)}i.
It should be noted that the quantity Mi of quantization intervals in dimension i is a function of the quantization cell selected in dimensions i+1, . . . ,D−1 and, therefore, can be calculated iteratively starting from MD−1. It is not possible to calculate in advance.
The assignment of an index Nε{0,1, . . . ,MD−1} to the current quantization cell and the reconstruction can be carried out using linked lookup tables as is known from shell mapping, e.g., according to Robert F. H. Fischer, Precoding and Signal Shaping for Digital Transmission, pp. 258-281, John Wiley & Sons, Inc., New York, 2002, ISBN 0471 22410 3:
The assignment of an index to the radius is not dependent upon the cell index on the unit sphere so that the problem of index assignment for the surface of the unit sphere can now be considered.
The starting number of the index is 0, the maximum ending number of the index is Mφ−1 corresponding to the quantity of quantization cells on the surface of the unit sphere. MD−1 is the quantity of subspheres of dimensionality D−1 and can be calculated from (23). A first-order lookup table with MD−1 entries N0, . . . , NM
Accordingly, there are Ni+1−Ni cells for a fixed ith value of {circumflex over (φ)}D−1. For every Ni, a second-order lookup table is applied which contains the indices Oij which again designate the smallest index of all of the cells of the jth quantization interval for the angle {circumflex over (φ)}D−2 within the D−1-dimensional subsphere to which Ni refers.
This procedure leads to D−2 orders of linked lookup tables in order to indicate the D−1 angles (the D−1-order lookup table for angle {circumflex over (φ)}1 would contain consecutive cell indices and must therefore not be tabularized).
The quantization noise and the signal/noise ratio are considered at greater length in the following.
As long as there is a sufficiently large quantity of quantization cells in D dimensions, the usual approximation of an equidistributed quantization error can be applied within these cubic cells, each cell being represented by its center. The output of the quantization noise is given approximately by:
The spherical logarithmic quantization forces Δ(r)=Δ·r in all dimensions for
Therefore,
see also (22). As is wanted, the SNR is independent from the variance of the signal in this range.
Inserting (13) and (16) gives:
When (29) is considered, F(D) can be interpreted as a loss with respect to the rate-distortion bound for iid Gaussian random variables (6 dB-per-bit rule). This relationship is shown in
Stirling's approximation
is used to calculate the limit
representing the loss in relation to the rate-distortion bound which is caused by the suboptimal cubic quantization cells (relative to (D−1)-dimensional hyperspheres for the quantization of the surface of a D-dimensional unit sphere and D→∞). In other words, spherical logarithmic quantization makes it possible to compensate again for the loss described in (1) resulting from companding to only 1.53 dB. Therefore, through the selection of the parameter A and D, it is possible in theory to achieve a dynamic range of any size (range of constant signal/noise ratio over the average signal level) without having to accept significant losses in the maximum signal/noise ratio that can be achieved. The asymptotic SNR loss of 1.53 dB corresponds to a rate loss of 1/4 bit/sample which is acceptable for the sake of an extreme reduction in complexity.
In view of the fact that no knowledge of a probability density function for iid source signal values can be deliberately assumed for the purpose of logarithmic companding, the rate-distortion function for the Gaussian random variables again gives the lower bound for the achievable rate (or distortion) at a given distortion (or rate) (Berger's upper bound of the rate-distortion function according to T. Berger, Lossy Source Coding, IEEE Transactions on Information Theory, pp. 2693-2723, October 1998). Therefore, improvements going beyond 1.53 dB or 1/4 bit/sample are not possible anyway under the present requirements and restrictions.
Further, it is clear from
BD≈B1+20 log10(r0)+10 log10(D). (33)
In addition, because of the averaging effect within D values, the saturation strength increases so that a further increase in the dynamic range is achieved. There is an infinitely large dynamic range for each desired value of A in limit (D→∞).
As in every vector quantization method, the delay of spherical logarithmic quantization by nature of its structure is equal to exactly D samples. As can be seen from
The combination of a spherical quantization and DPCM is described in the following.
Correlations between the samples q[k] are made use of efficiently by means of differential PCM (DPCM) (see
It should be noted that a gain in SNR expressed by a reduction in the mean square error between the original samples and the reconstructed samples is possible when using the so-called backward prediction shown in
With logarithmic quantization, the SNR of the prediction error signal x[k] is not dependent on its variance or, to express this differently, the output of the quantization noise is proportional to the signal output. Accordingly, the prediction gain, i.e., the quotient of the variances of q[k] and x[k] can be converted directly to a gain in SNR. Therefore, logarithmic quantization is a good choice for DPCM. In addition to this, no further signal delay inherent to the system is carried out by DPCM compared with PCM because an optimal prediction error filter for maximum prediction gain is fundamentally causal and strictly minimum phase (see L. Pakula and S. Kay, Simple Proofs of the Minimum Phase Property of the Prediction Error Filter, IEEE Transactions on ASSP, Vol. 31, 1983) and can therefore be inverted without structure-induced delay.
The gradient descent method will be described more fully in the following.
Application of spherical logarithmic quantization to DPCM with backward prediction is subject to the same set of problems as any other vector quantization method: in order to calculate the current prediction error signal x[k], all of the preceding reconstructed samples q[k−i], i=1, 2, . . . must be present. For a high prediction gain, the immediately preceding values (i=1, i=2) in particular are indispensable. Unfortunately, this requirement cannot be reconciled with the quantization of blocks of length D samples.
Any of the methods mentioned in the literature for combining vector quantization with DPCM can be used to solve this problem. In the present case, a method relying on the principle of analysis by synthesis is used. This method is known, for example from CELP waveform coding methods according to N. S. Jayant, P. Noll, Digital Coding of Waveforms, Prentice-Hall, Englewood Cliffs, N.J., 1984, and this approach is combined with a discrete gradient descent method.
In order to resolve the conflict between DPCM and vector quantization, the square Euclidean distance between vectors of samples
q[l]=(q[l·D],q[l·D+1], . . . ,q[l·D+D−1]) (34)
and a corresponding reconstruction vector {circumflex over (q)}[l] is minimized l=└k/D┘.
It should be noted at this point that, in addition to spherical logarithmic quantization, the calculation of the corresponding prediction error signal and the inversion of the prediction error filter must be included in the calculation of a pair q,{circumflex over (q)}, wherein preceding reconstruction vectors {circumflex over (q)}[l−m],m=1,2, . . . are resorted to, but themselves remain unchanged. The aim of the algorithm is to find those quantization cells for x[l] for which the metric
is minimized.
In order to find a suitable starting value for the algorithm, we suggest beginning with a forward prediction for the current D samples corresponding to a deactivation of the chain from ADC and DAC in
Since linear prediction is included in the optimizing process, the resulting SNR is often greater than can be predicted by adding the gains (in dB) from spherical logarithmic quantization and prediction (DPCM). Noteworthy gains in SNR can be observed particularly in case of small rates (e.g., R<5 bits/sample), which will be described in the following and with reference to
Simulations show that the required number of iterations is on average approximately 0.25 (per D sample) and, when limited to a maximum of 3 iterations, no significant losses are observed compared to an unlimited search space, so that the search for the optimal quantization cell for low dimension numbers can certainly be implemented in real time. Further, small values of D already yield large gains (see
On the reception side, there is no change compared with conventional DPCM methods. It should be noted here that the total delay of the transmission system is only D samples and it is excellently suited to transmissions which must meet the requirements of an extremely short delay.
Variants of the method are described more fully in the following.
To speed up the search for the most favorable quantization leading to minimum distortion, e.g., according to (35), all algorithms of the lattice decoding are applicable in principle or for finding a maximum likelihood codeword in channel coding (see, for example, Erik Agrell, Thomas Eriksson, Alexander Vardy, Kenneth Zeger, Closest Point Search in Lattices, IEEE Transactions on Information Theory, pp. 2201-2214, August 2002, and the references cited therein). This method is preferably transformed into spherical coordinates.
A variant without iterative determination of the quantization cell is provided in that a non-uniform quantization of the angle coordinates is carried out instead of taking into account statistical dependencies within the D signal values which are actually to be quantized by means of a linear prediction filter according to
It is suggested that the logarithmic quantization of the amount is maintained so that the signal points can be normalized subsequently in such a way that they lie on a sphere with radius 1. This allows the analytic calculation of the probability density function of the signal points on the spherical surface, e.g., assuming a Gaussian signal process from the autocorrelation function of the source signal or a direct empirical determination of relative occurrences of the signal points on the spherical surface.
A non-uniform cubic quantization of the surface of a sphere with radius 1 can be determined in this way for M>>1, e.g., by means of the optimization formula according to (9). This formula states that, on the average, the amount of every quantization should be equal to the quantization noise. When Δi is the edge length of a (D−1) -dimensional cubic quantization cell and zi is the associated center point, that is, the reception-side reconstruction vector with amount 1, and ƒ(v);v:=(φ1, . . . ,φD−1) is the probability density function or relative occurrence of signal values on the spherical surface, the following equation should be approximately satisfied:
with the secondary condition
This leads directly to a (D−1)-dimensional compressor function k(v) by which the non-uniform quantization of the sphere surface is uniquely determined. The non-uniform quantization can then be completed, for example, as in the one-dimensional case, by a nonlinear deformation of vector v to a vector z:=k(v), subsequent uniform quantization to vector zi according to III-C, and application of the inverse function {circumflex over (v)}:=k−1 (zi) subsequently for obtaining the reconstruction vector.
The resulting (D−1)-dimensional compressor function should preferably be approximated by an analytically describable function or, analogous to the usual representation in one-dimensional compression by means of straight-line segments (e.g., 13-segment characteristic line, see [5]), by (D−2)-dimensional partial planes (with constant partial derivations). When this approximation is latticed orthogonally over the space
over the D−1 angles, the compressor function can be evaluated in a simple manner.
To simulate spherical logarithmic quantization in combination with DPCM, the quantization of the overture and the aria “Der Vogelfänger bin ich ja” from the opera “Zauberflöte” by Wolfgang Amadeus Mozart was examined (Philips Classics Productions 1994 (DDD), Mozart, “Der Vogelfänger bin ich ja” (Die Zauberflöte), Polygram Records #442569-2, track 3).
The signal was first coded, then decoded and the SNR was calculated by comparison with the original CD signal. The signal/noise ratio is averaged over the entire piece of music. At the low rates, the SNR rises more steeply with D compared to
The average signal level of the overture is around −27.20 dB; this piece of music is characterized by a very large dynamic range from below −70 dB to −17 dB (see also
The value A=48270 corresponds to a compression of 12 bits, e.g., for D=1, R=12 a resolution of 24 bits is achieved for small signal values. Of course, there is never a resolution corresponding to 16 bits for D=1, R=4 because the interval number is too low. However, when spherical logarithmic quantization is applied in 6 dimensions, the extremely large value of A=48270 is actually well chosen for a high minimum value of the SNR. In this case, the resolution is even identical to that of the original CD data for segments with a signal level of −70 dB and lower. To date, no one has been found, even among trained personnel, who was capable of reliably hearing a difference between the quantized signal at R=4, D=6, A=48270 and the original signal.
The preceding description shows a waveform-preserving digitization method for analog source signals which, on the one hand, combines the gain by multidimensional quantization with the advantages of logarithmic quantization and, on the other hand, is capable of additionally increasing the objective signal/noise ratio through the addition of prediction gains. In conclusion, it is noted that, in addition to a favorable exchange of rate and distortion, this method is distinguished above all by an extremely large dynamic range accompanied at the same time by an extremely low system-induced delay of the signal by only a few sample intervals.
While the foregoing description and drawings represent the present invention, it will be obvious to those skilled in the art that various changes may be made therein without departing from the true spirit and scope of the present invention.
Number | Date | Country | Kind |
---|---|---|---|
103 34 357 | Jul 2003 | DE | national |
Filing Document | Filing Date | Country | Kind | 371c Date |
---|---|---|---|---|
PCT/EP2004/008217 | 7/23/2004 | WO | 00 | 2/12/2007 |
Publishing Document | Publishing Date | Country | Kind |
---|---|---|---|
WO2005/013492 | 2/10/2005 | WO | A |
Number | Name | Date | Kind |
---|---|---|---|
4433604 | Ott | Feb 1984 | A |
6028634 | Yamaguchi et al. | Feb 2000 | A |
6198830 | Holube et al. | Mar 2001 | B1 |
20020176353 | Atlas et al. | Nov 2002 | A1 |
Number | Date | Country |
---|---|---|
0 360 770 | Mar 1990 | EP |
Number | Date | Country | |
---|---|---|---|
20070211804 A1 | Sep 2007 | US |