Binaural multi-channel decoder in the context of non-energy-conserving upmix rules

Description of equivalent WO 2007140809 A1

0001Binaural Multi-Channel Decoder In The Context Of Non- Energy-Conserving Upmix Rules

0002Field of the Invention

0003The present invention relates to binaural decoding of multi-channel audio signals based on available downmixed signals and additional control data, by means of HRTF filtering.

0004Background of the Invention and Prior Art

0005Recent development in audio coding has made methods available to recreate a multi-channel representation of an audio signal based on a stereo (or mono) signal and corresponding control data. These methods differ substantially from older matrix based solution such as Dolby Prologic, since additional control data is transmitted to control the re-creation, also referred to as up-mix, of the surround channels based on the transmitted mono or stereo channels.

0006Hence, such a parametric multi-channel audio decoder, e.g. MPEG Surround reconstructs N channels based on M transmitted channels, where N>M , and the additional control data. The additional control data represents a significantly lower data rate than that required for transmission of all N channels, making the coding very efficient while at the same time ensuring compatibility with both M channel devices and N channel devices. [J. Breebaart et al. "MPEG spatial audio coding / MPEG Surround: overview and current status", Proc. 119th AES convention, New York, USA, October 2005, Preprint 6447] .

0007These parametric surround coding methods usually comprise a parameterization of the surround signal based on Channel Level Difference (CLD) and Inter-channel coherence/cross- correlation (ICC) . These parameters describe power ratios and correlation between channel pairs in the up-mix process. Further Channel Prediction Coefficients (CPC) are also used in prior art to predict intermediate or output channels during the up-mix procedure.

0008Other developments in audio coding have provided means to obtain a multi-channel signal impression over stereo headphones. This is commonly done by downmixing a multichannel signal to stereo using the original multi-channel signal and HRTF (Head Related Transfer Functions) filters.

0009Alternatively, it would, of course, be useful for computational efficiency reasons and also for audio quality reasons to short-cut the generation of the binaural signal having the left binaural channel and the right binaural channel .

0010However, the question is how the original HRTF filters can be combined. Further a problem arises in a context of an energy-loss-affected upmixing rule, i.e., when the multichannel decoder input signal includes a downmix signal having, for example, a first downmix channel and a second downmix channel, and further having spatial parameters, which are used for upmixing in a non-energy-conserving way. Such parameters are also known as prediction parameters or CPC parameters. These parameters have, in contrast to channel level difference parameters the property that they are not calculated to reflect the energy distribution between two channels, but they are calculated for performing a best-as-possible waveform matching which automatically results in an energy error (e.g. loss), since, when the prediction parameters are generated, one does not care about energy-conserving properties of an upmix, but one does care about having a good as possible time or subband domain waveform matching of the reconstructed signal compared to the original signal. When one would simply linearly combine HRTF filters based on such transmitted spatial prediction parameters, one will receive artifacts which are especially serious, when the prediction of the channels performs poorly. In that situation, even subtle linear dependencies lead to undesired spectral coloring of the binaural output. It has been found out that this artifact occurs most frequently when the original channels carry signals that are pairwise uncorrelated and have comparable magnitudes.

0011Summary of the Invention

0012It is the object of the present invention to provide an efficient and qualitatively acceptable concept for multichannel decoding to obtain a binaural signal which can be used, for example, for headphone reproduction of a multichannel signal.

0013In accordance with the first aspect of the present invention, this object is achieved by a multi-channel decoder for generating a binaural signal from a downmix signal derived from an original multi-channel signal using parameters including an upmix rule information useable for upmixing the downmix signal with an upmix rule, the upmix rule resulting in an energy-error, comprising: a gain factor calculator for calculating at least one gain factor for reducing or eliminating the energy-error, based on the upmix rule information and filter characteristics of a head related transfer function based filters corresponding to upmix channels, and a filter processor for filtering the downmix signal using the at least one gain factor, the filter characteristics and the upmix rule information to obtain an energy-corrected binaural signal.

0014In accordance with a second aspect of this invention, this object is achieved by a method of multi-channel decoding for generating a binaural signal from a downmix signal derived from an original multi-channel signal using parameters including an upmix rule information useable for upmixing the downmix signal with an upmix rule, the upmix rule resulting in an energy-error, comprising: calculating at least one gain factor for reducing or eliminating the energy-error, based on the upmix rule information and filter characteristics of a head related transfer function based filters corresponding to upmix channels; and filtering the downmix signal using the at least one gain factor, the filter characteristics and the upmix rule information to obtain an energy-corrected binaural signal.

0015Further aspects of this invention relate to a computer program having a computer-readable code which implements, when running on a computer, the method of multi-channel decoding.

0016The present invention is based on the finding that one can even advantageously use up-mix rule information on an upmix resulting in an energy error for filtering a downmix signal to obtain a binaural signal without having to fully render the multichannel signal and to subsequently apply a huge number of HRTF filters. Instead, in accordance with the present invention, the upmix rule information relating to an energy-error-affected upmix rule can advantageously be used for short-cutting binaural rendering of a downmix signal, when, in accordance with the present invention, a gain factor is calculated and used when filtering the downmix signal, wherein this gain factor is calculated such that the energy error is reduced or completely eliminated.

0017Particularly, the gain factor not only depends on the information on the upmix rule such as the prediction parameters, but, importantly, also depends on head related transfer function based filters corresponding to upmix channels, for which the upmix rule is given. Particularly, these upmix channels never exist in the preferred embodiment of the present invention, since the binaural channels are calculated without firstly rendering, for example, three intermediate channels. However, one can derive or provide HRTF based filters corresponding to the upmix channels although the upmix channels themselves never exist in the preferred embodiment. It has been found out that the energy error introduced by such an energy-loss- affected upmix rule not only corresponds to the upmix rule information which is transmitted from the encoder to the decoder, but also depends on the HRTF based filters so that, when generating the gain factor, the HRTF based filters also influence the calculation of the gain factor.

0018In view of that, the present invention accounts for the interdependence between upmix rule information such as prediction parameters and the specific appearance of the HRTF based filters for the channels which would be the result of upmixing using the upmix rule.

0019Thus, the present invention provides a solution to the problem of spectral coloring arising from the usage of a predictive upmix in combination with binaural decoding of parametric multi-channel audio.

0020Preferred embodiments of the present invention comprise the following features: an audio decoder for generating a binaural audio signal from M decoded signals and spatial parameters pertinent to the creation of N>M channels, the decoder comprising a gain calculator for estimating, in a multitude of subbands, two compensation gains from P pairs of binaural subband filters and a subset of the spatial parameters pertinent to the creation of P intermediate channels, and a gain adjuster for modifying, in a multitude of subbands, M pairs of binaural subband filters obtained by linear combination of the P pairs of binaural subband filters, the modification consisting of multiplying each of the M pairs with the two gains computed by the gain calculator. Brief Description of the Drawings

0021The present invention will now be described by way of illustrative examples, not limiting the scope or spirit of the invention, with reference to the accompanying drawings, in which:

0022Fig. 1 illustrates binaural synthesis of parametric multichannel signals using HRTF related filters;

0023Fig. 2 illustrates binaural synthesis of parametric multichannel signals using combined filtering;

0024Fig. 3 illustrates the components of the inventive parameter/filter combiner;

0025Fig. 4 illustrates the structure of MPEG Surround spatial decoding;

0026Fig. 5 illustrates the spectrum of a decoded binaural signal without the inventive gain compensation;

0027Fig. 6 illustrates the spectrum of the inventive decoding of a binaural signal.

0028Fig. 7 illustrates a conventional binaural synthesis using HRTFs;

0029Fig. 8 illustrates a MPEG surround encoder;

0030Fig. 9 illustrates cascade of MPEG surround decoder and binaural synthesizer;

0031Fig. 10 illustrates a conceptual 3D binaural decoder for certain configurations; Fig. 11 illustrates a spatial encoder for certain configurations;

0032Fig. 12 illustrates a spatial (MPEG Surround) decoder;

0033Fig. 13 illustrates filtering of two downmix channels using four filters to obtain binaural signals without gain factor correction;

0034Fig. 14 illustrates a spatial setup for explaining different HRTF filters 1-10 in a five channels setup;

0035Fig. 15 illustrates a situation of Fig. 14, when the channels for L, Ls and R, Rs have been combined;

0036Fig. 16a illustrates the setup from Fig. 14 or Fig. 15, when a maximum combination of HRTF filters has been performed and only the four filters of Fig. 13 remain;

0037Fig. 16b illustrates an upmix rule as determined by the Fig. 20 encoder having upmix coefficients resulting in a non-energy-conserving upmix;

0038Fig. 17 illustrates how HRTF filters are combined to finally obtain four HRTF-based filters;

0039Fig. 18 illustrates a preferred embodiment of an inventive multi-channel decoder;

0040Fig. 19a illustrates a first embodiment of the inventive multi-channel decoder having a scaling stage after HRTF-based filtering without gain correction; Fig. 19b illustrates an inventive device having adjusted HRTF-based filters which result in a gain- adjusted filter output signal; and

0041Fig. 20 shows an example for an encoder generating the information for a non-energy-conserving upmix rule .

0042Detailed description of preferred embodiments

0043Before discussing the inventive gain adjusting aspect in detail, a combination of HRTF filters and usage of HRTF- based filters will be discussed in connection with Figs. 7 to 11.

0044In order to better outline the features and advantages of the present invention a more elaborate description is given first. A binaural synthesis algorithm is outlined in Fig. 7. A set of input channels is filtered by a set of HRTFs. Each input signal is split in two signals (a left ^ΛL' , and a right ^XR' component) ; each of these signals is subsequently filtered by an HRTF corresponding to the desired sound source position. All left-ear signals are subsequently summed to generate the left binaural output signal, and the right-ear signals are summed to generate the right binaural output signal.

0045The HRTF convolution can be performed in the time domain, but it is often preferred to perform the filtering in the frequency domain due to computational efficiency. In that case, the summation as shown in Fig. 7 is also performed in the frequency domain.

0046In principle, the binaural synthesis method as outlined in Fig. 7 could be directly used in combination with an MPEG surround encoder/decoder. The MPEG surround encoder is schematically shown in Fig. 8. A multi-channel input signal is analyzed by a spatial encoder, resulting in a mono or stereo down mix signal, combined with spatial parameters. The down mix can be encoded with any conventional mono or stereo audio codec. The resulting down-mix bit stream is combined with the spatial parameters by a multiplexer, resulting in the total output bit stream.

0047A binaural synthesis scheme in combination with an MPEG surround decoder is shown in Fig. 9. The input bit stream is de-multiplexed resulting in spatial parameters and a down-mix bit stream. The latter bit stream is decoded using a conventional mono or stereo decoder. The decoded down mix is decoded by a spatial decoder, which generates a multichannel output based on the transmitted spatial parameters. Finally, the multi-channel output is processed by a binaural synthesis stage as depicted in Fig. 7, resulting in a binaural output signal.

0048There are however at least three disadvantages of such a cascade of an MPEG surround decoder and a binaural synthesis module:

0049• A multi-channel signal representation is computed as an intermediate step, followed by HRTF convolution and downmixing in the binaural synthesis step. Although HRTF convolution should be performed on a per channel basis, given the fact that each audio channel can have a different spatial position, this is an undesirable situation from a complexity point of view.

0050• The spatial decoder operates in a filterbank (QMF) domain. HRTF convolution, on the other hand, is typically applied in the FFT domain. Therefore, a cascade of a multi-channel QMF synthesis filterbank, a multi-channel DFT transform, and a stereo inverse DFT transform is necessary, resulting in a system with high computational demands. • Coding artifacts created by the spatial decoder to create a multi-channel reconstruction will be audible, and possibly enhanced in the (stereo) binaural output.

0051The spatial encoder is shown in Fig. 11. A multi-channel input signal consisting of Lf, Ls, C, Rf and Rs signals, for the left-front, left-surround, center, right-front and right-surround channels is processed by two ^λOTT" units, which both generate a mono down mix and parameters for two input signals. The resulting down-mix signals, combined with the center channel are further processed by a ^TTT' (Two-To-Three) encoder, generating a stereo down mix and additional spatial parameters.

0052The parameters resulting from the ^TTT' encoder typically consist of a pair of prediction coefficients for each parameter band, or a pair of level differences to describe the energy ratios of the three input signals. The parameters of the *OTT' encoders consist of level differences and coherence or cross-correlation values between the input signals for each frequency band.

0053In Fig. 12 a MPEG Surround decoder is depicted. The downmix signals 10 and r0 are input into a Two-To-Three module, that recreates a center channel, a right side channel and a left side channel. These three channels are further processed by several OTT modules (One-To-Two) yielding the six output channels.

0054The corresponding binaural decoder as seen from a conceptual point of view is shown in Fig. 10. Within the filterbank domain, the stereo input signal (L₀, Ro) is processed by a TTT decoder, resulting in three signals L, R and C. These three signals are subject to HRTF parameter processing. The resulting 6 channels are summed to generate the stereo binaural output pair {L_b, Rb) . The TTT decoder can be described as the following matrix operation:

0055<img file="WO2007140809A1_D0001.tif" />

0056with matrix entries m_xy dependent on the spatial parameters. The relation of spatial parameters and matrix entries is identical to those relations as in the 5.1- multichannel MPEG surround decoder. Each of the three resulting signals L, R, and C are split in two and processed with HRTF parameters corresponding to the desired

0057(perceived) position of these sound sources. For the center channel (C) , the spatial parameters of the sound source position can be applied directly, resulting in two output signals for center, L₈(C) and R₃(C):

0058<img file="WO2007140809A1_D0002.tif" />

0059For the left (L) channel, the HRTF parameters from the left-front and left-surround channels are combined into a single HRTF parameter set, using the weights w_lf and w_rf .

0060The resulting Composite' HRTF parameters simulate the effect of both the front and surround channels in a statistical sense. The following equations are used to generate the binaural output pair (L_B, R_B) for the left channel :

0061<img file="WO2007140809A1_D0003.tif" />

0062In a similar fashion, the binaural output for the right channel is obtained according to: L₈(R) H_L(R)

R

0064H_R⁽R)

0065Given the above definitions of L₈(C), R₈(C), L₈(L), R₈(L), L_B(R) and R₃(R), the complete L_B and R₈ signals can be derived from a single 2 by 2 matrix given the stereo input signal:

0066<img file="WO2007140809A1_D0004.tif" />

0067with

0068h_u=m_nH_L(L) + m_2lH_L(R) + m_ixH_L(C), <img file="WO2007140809A1_D0005.tif" /> K = m_nH_R(L) + m_2lH_R(R) + m_3lH_R(C) , h₂₂=m_l2H_Ji(L) + m₂₂H_R(R) + m_i2H_R(C) .

0069The Hx(Y) filters can be expressed as parametric weighted combinations of parametric versions of the original HRTF filters. In order for this to work, the original HRTF filters are expressed as a

0070• An (average) level per frequency band for the left-ear impulse response;

0071• An (average) level per frequency band for the right- ear impulse response;

0072• An (average) arrival time or phase difference between the left-ear and right-ear impulse response.

0073Hence, the HRTF filters for the left and right ear given the center channel input signal is expressed as: H_L{C) P₁(W jΦiOn P_r(C)e -JΦ(C)I2

0074where P₁(C) is the average level for a given frequency band for the left ear, and «αis the phase difference.

0075Hence, the HRTF parameter processing simply consists of a multiplication of the signal with P_x and P_r corresponding to the sound source position of the center channel, while the phase difference is distributed symmetrically. This process is performed independently for each QMF band, using the mapping from HRTF parameters to QMF filterbank on the one hand, and mapping from spatial parameters to QMF band on the other hand.

0076Similarly the HRTF filters for the left and right ear given the left channel and right channel are given by:

0077H_L(L) = ,JwJP₁² (Lf) + wf_sP,² (Ls) ,

0078<img file="WO2007140809A1_D0006.tif" />

0079H_L(R) = e⁺'W*'-""* ^P² (Rf) + w²_sP² (Rs) ,

0080<img file="WO2007140809A1_D0007.tif" />

0081Clearly, the HRTFs are weighted combinations of the levels and phase differences for the parameterized HRTF filters for the six original channels.

0082The weights w_lf and w_ls depend on the CLD parameter of the ^ΛOTT' box for Lf and Ls:

0083/10 ' '* _{1 + 1} QCLD, /10 ' <img file="WO2007140809A1_D0008.tif" /> And the weights w_rf and w_rs depend on the CLD parameter of the ^ΛOTT' box for Rf and Rs:

0084' <img file="WO2007140809A1_D0009.tif" /> ' "r "i ι ₊ i QCLD_r no ^•

0085The above approach works well for short HRTF filters that sufficiently accurate can be expressed as an average level per frequency band, and an average phase difference per frequency band. However, for long echoic HRTFs this is not the case.

0086The present invention teaches how to extend the approach of a 2 by 2 matrix binaural decoder to handle arbitrary length HRTF filters. In order to achieve this, the present invention comprises the following steps:

0087• Transform the HRTF filter responses to a filterbank domain;

0088• Overall delay difference or phase difference extraction from HRTF filter pairs;

0089• Morph the responses of the HRTF filter pair as a function of the CLD parameters

0090• Gain adjustment

0091This is achieved by replacing the six complex gains H_Y{X) for 7 = Z₀,i?_oand X = L,R,C with six filters. These filters are derived from the ten filters H_Y(X) for Y = L₀,R₀ and X = Lf,Ls,Rf,Rs,C , which describe the given HRTF filter responses in the QMF domain. These QMF representations can be achieved according to the method described below.

0092The morphing of the front and surround channel filters is performed with a complex linear combination according to

0093H_γ{X)^gw_f^{,-jφ_xγw])H_ϊi_&lt;Xf)+gw_s^{jφ_χγw))H_γ{X_S) . The phase parameter φ^ can be defined from the main delay time difference ^between the front and back HRTF filters and the subband index not the QMF bank via

0094JΓ(;»₊^l)

0095Ψxr ₆₄^τxr >

0096The role of this phase parameter in the morphing of filters is twofold. First, it realizes a delay compensation of the two filters prior to superposition which leads to a combined response which models a main delay time corresponding to a source position between the front and the back speakers. Second, it makes the necessary gain compensation factor gmuch more stable and slowly varying over frequency than in the case of simple superposition with φ_xγ=0.

0097The gain factor gis determined by the same incoherent addition power rule as for the parametric HRTF case,

0098Py(Xf = W]Py (Xf f + W* Py (Xsf ,

0099where

0100Py (Xf = g² (rfPy (Xf f + ™]Py (Xsf + 2w_fW_sP_γ {Xf)P_γ {Xs)_Pχγ )

0101andp_χy is the real value of the normalized complex cross correlation between the filters

0102<img file="WO2007140809A1_D0010.tif" />

0103In the case of simple superposition with^_ΛΥ=0_/ the value of P_1n varies in an erratic and oscillatory manner as a function of frequency, which leads to the need for extensive gain adjustment. In practical implementation it is necessary to limit the value of the gaingand a remaining spectral colorization of the signal cannot be avoided.

0104In contrast, the use of morphing with a delay based phase compensation as taught by the present invention leads to a smooth behavior of Pχγ as a function of frequency. This value is often even close to one for natural HRTF derived filter pairs since they differ mainly in a delay and amplitude, and the purpose of the phase parameter is to take the delay difference into account in the QMF filterbank domain.

0105An alternative beneficial choice of phase parameter ^ is given by computing the phase angle of the normalized complex cross correlation between the filters HAXf) _and H_y(Xs) _f and unwrapping the phase values with standard unwrapping techniques as a function of the subband index "of the QMF

0106bank. This choice has the consequence that ^-^is never negative and hence the compensation gain £ satisfies

0107^~£^~ for all subbands . Moreover this choice of phase parameter enables the morphing of the front and surround channel filters in situations where a main delay time

0108difference '"'is not available.

0109All signals considered below are subband samples from a modulated filter bank or windowed FFT analysis of discrete time signals or discrete time signals. It is understood that these subbands have to be transformed back to the discrete time domain by corresponding synthesis filter bank operations.

0110Fig. 1 illustrates a procedure for binaural synthesis of parametric multichannel signals using HRTF related filters. A multichannel signal comprising N channels is produced by spatial decoding 101 based on M &lt; N transmitted channels and transmitted spatial parameters. These N channels are in turn converted into two output channels intended for binaural listening by means of HRTF filtering. This HRTF filtering 102 superimposes the results of filtering each input channel with one HRTF filter for the left ear and one HRTF filter for the right ear. All in all, this requires 2N filters. Whereas the parametric multichannel signal achieves a high quality listener experience when listened to through N loudspeakers, subtle interdependencies of the N signals will lead to artifacts for the binaural listening. These artifacts are dominated by deviation in spectral content from the reference binaural signal as defined by HRTF filtering of the original N channels prior to coding. A further disadvantage of this concatenation is that the total computational cost for binaural synthesis is the addition of the cost required for each of the components 101 and 102.

0111Fig. 2 illustrates binaural synthesis of parametric multichannel signals by using the combined filtering taught by the present invention. The transmitted spatial parameters are split by 201 into two sets, Set 1 and Set 2. Here, Set 2 comprises parameters pertinent to the creation of P intermediate channels from the M transmitted channels and Set 1 comprises parameters pertinent to the creation of //channels from theP intermediate channels. The prior art precombiner 202 combines selected pairs of the 2N HRTF related subband filters with weights that depend the parameter Set 1 and the selected pairs of filters. The result of this precombination is 2P binaural subband filters which represent a binaural filter pair for each of the P intermediate channels. The inventive combiner 203 combines the 2P binaural subband filters into a set of 2M binaural subband filters by applying weights that depend both on the parameter Set 2 and the 2P binaural subband filters. In comparison, a prior art linear combiner would apply weights that depend only on the parameter Set 2. The resulting set of 2M filters consists of a binaural filter pair for each of the M transmitted channels. The combined filtering unit 204 obtains a pair of contributions to the two channel output for each of the M transmitted channels by filtering with the corresponding filter pair. Subsequently, all the M contributions are added up to form a two channel output in the subband domain.

0112Fig. 3 illustrates the components of the inventive combiner 203 for combination of spatial parameters and binaural filters. The linear combiner 301 combines the IP binaural subband filters into 2M binaural filters by applying weights that are derived from the given spatial parameters, where these spatial parameters are pertinent to the creation of P intermediate channels from the M transmitted channels. Specifically, this linear combination simulates the concatenation of an upmix from M transmitted channels to P intermediate channels followed by a binaural filtering from P sources. The gain adjuster 303 modifies the IM binaural filters output from the linear combiner 301 by applying a common left gain to each of the filters that correspond to the left ear output and by applying a common right gain to each of the filters that correspond to the right ear output. Those gains are obtained from gain calculator 302 which derives the gains from the spatial parameters and the 2P binaural filters. The purpose of the gain adjustment of the inventive components 302 and 303 is to compensate for the situation where the P intermediate channels of the spatial decoding carry linear dependencies that lead to unwanted spectral coloring due to the linear combiner 301. The gain calculator 302 taught by the present invention includes means for estimating an energy distribution of the P intermediate channels as a function of the spatial parameters.

0113Fig. 4 illustrates the structure of MPEG Surround spatial decoding in the case of a stereo transmitted signal. The analysis subbands of the M - 2 transmitted signals are fed into the 2—»3 box 401 which outputs P = 3 intermediate signals, a combined left, a combined right, and a combined center. This upmix depends on a subset of the transmitted spatial parameters which corresponds to Set 2 on Fig. 2. The three intermediate signals are subsequently fed into three 1—»2 boxes 402-404 which generate a totality ofN = 6 signals 405: l_f (left front), Z₅ (left surround), r_f (right front) , r_s (right surround) , c (center) , and lfe (low frequency extension) . This upmix depends on a subset of the transmitted spatial parameters which corresponds to Set 1 on Fig. 2. The final multichannel digital audio output is created by passing the six subband signals into six synthesis filter banks.

0114Fig. 5 illustrates the problem to be solved by the inventive gain compensation. The spectrum of a reference

0115HRTF filtered binaural output for the left ear is depicted as a solid graph. The dashed graph depicts the spectrum of the corresponding decoded signal as generated by the method of Fig. 2, in the case where the combiner 203 consists of the linear combiner 301 only. As it can be seen, there is a substantial spectral energy loss relative to the desired reference spectrum in the frequency intervals 3-4 kHz and 11-13 kHz. There is also a smaller spectral boost around 1 kHz and 10 kHz.

0116Fig. 6 illustrates the benefit of using the inventive gain compensation. The solid graph is the same reference spectrum as in Fig. 5, but now the dashed graph depicts the spectrum of the decoded signal as generated by the method of Fig. 2, in the case where the combiner 203 consists of all the components of Fig. 3. As it can be seen, there is a significantly improved spectral match between the two curves compared to that of the two curves of Fig. 5.

0117In the text which follows, the mathematical description of the inventive gain compensation will be outlined. For discrete complex signals x,y , the complex inner product and squared norm (energy) is defined by

0118<img file="WO2007140809A1_D0011.tif" />

0119where y{k) denotes the complex conjugate signal of y(k) .

0120The original multichannel signal consists of N channels, and each channel has a binaural HRTF related filter pair associated to it. It will however be assumed here that the parametric multichannel signal is created with an intermediate step of predictive upmix from the M transmitted channels to P predicted channels. This structure is used in MPEG Surround as described by Fig. 4. It will be assumed that the original set of IN HRTF related filters have been reduced by the prior art precombiner 202 to a filter pair for each of the P predicted channels where M ≤P≤N . The P predicted channel signals x_p , p = l,2,...,P , aim at approximating theP signals x_p , p = \,2,...,P , which are derived from the original N channels via partial downmix. In MPEG Surround, these signals are a combined left, a combined right and a combined and scaled center/Ife channel. It is assumed that the HRTF filter pair corresponding to the signal x_p is described by a subband filter b_ip for the left ear and a subband filter b_2pfor the right ear. The reference binaural output signal is thus given by the linear superposition of filtered signals for « = 1,2,

0121y_n(k)=∑(b_n,p-x_p)(k)_f

0122P=I

0123where the star denotes convolution in the time direction. The subband filters can be given in form of finite impulse response (FIR) filters, infinite impulse response (HR) or derived from a parameterized family of filters.

0124In the encoder, the downmix is formed by the application of a MxP downmix matrix D to a column vector of signals formed by x_p p = l,2,...,P and the prediction in the decoder is performed by the application of a PxM prediction matrix C to the column vector of signals formed by the M transmitted downmixed channels z_m m = \,...,M ,

0125<img file="WO2007140809A1_D0012.tif" />

0126Both matrices are known at the decoder, and ignoring the effects of coding the downmixed channels, the combined effect of prediction can be modeled by

0127<img file="WO2007140809A1_D0013.tif" />

0128where α P.R are the entries of the matrix product A= CD.

0129A straightforward method for producing a binaural output at the decoder is to simply insert the predicted signals X_p in (2) resulting in <img file="WO2007140809A1_D0014.tif" />

0130In terms of computations, the binaural filtering is combined with the predictive upmix beforehand such that (5) can be written as

0131<img file="WO2007140809A1_D0015.tif" />

0132with the combined filters defined by

0133<img file="WO2007140809A1_D0016.tif" />

0134This formula describes the action of the linear combiner 301 which combines the coefficients c_pm derived from spatial parameters with the binaural subband domain filters b_πp. When the original? signals*^ have a numerical rank essentially bounded by M , the prediction can be designed to perform very well and the approximation x_p∞x_p±s valid. This happens for instance if only M of the P channels are active, or if important signal components originate from amplitude panning. In that case the decoded binaural signal (5) is a very good match to the reference (2) . On the other hand, in the general case and especially in case the original P signals x_p are uncorrelated, there will be a substantial prediction loss and the output from (5) can have an energy that deviates considerably from the energy of (2). As the deviation will be different in different frequency bands, the final audio output suffers from spectral coloring artifacts as described by Fig. 5. The present invention teaches how to circumvent this problem by gain compensating the output according to y_n = g_n-y_n- (8)

0135In terms of computations, the gain compensation is advantageously performed by altering the combined filters according to the gain adjuster 303 , h_nm(k) = g_nh_nm(k) . The modified combined filtering then becomes

0136<img file="WO2007140809A1_D0017.tif" />

0137The optimal values of the compensating gains in (8) are

0138<img file="WO2007140809A1_D0018.tif" />

0139The purpose of the gain calculator 302 is to estimate these gains from the information available in the decoder.

0140Several tools for this end will now be outlined. The available information is represented here by the matrix entries a_pq and the HRTF related subband filters b_np .

0141First, the following approximation will be assumed for the inner product between signals x,ythat have been filtered by HRTF related subband filters b,d ,

0142{b*x,d_*y)_*{b,d){x,y). (11)

0143This approximation relies on the fact that often most energy of the filters is concentrated in a dominant single tap, which in turn presupposes that the time step of the applied time frequency transform is sufficiently large in comparison to the main delay differences of HRTF filters. Applying the approximation (11) in combination with (2) leads to kl²" Σ(*■>.,)(*,.*,>^• (12)

0144/>.Η

0145The next approximation consists of assuming that the original signals are uncorrelated, that is (^χ_p>^x_q) ⁼ 0 f°^r p≠q . Then (12) reduces to

0146lkf ^«Σ|U²lkf :i3)

0147For the decoded energy the result corresponding to (12) is

0148lΛf*Σ⁽*.^,A_Λ⁾⁽v*^,)- (14)

0149P,&lt;H

0150Inserting the predicted signals (4) in (14) and applying the assumption that the original signals are uncorrelated gives

0151<img file="WO2007140809A1_D0019.tif" />

0152What remains in order to be able to calculate the compensation gain given by the quotient (10) is to estimate the energy distribution Lc I , p = \,2,...,P of the original channels up to an arbitrary factor. The present invention teaches to do this by computing, as a function of the energy distribution, the prediction matrix C_model corresponding to the assumption that these channels are uncorrelated and that the encoder aims at minimizing the prediction error. The energy distribution is then estimated by solving the nonlinear system of equations C_model = C if possible. For prediction parameters that lead to a system of equations without solutions, the gain compensation factors are set to #„=1. This inventive procedure will be detailed in the following section in the most important special case.

0153The computation load imposed by (15) can be reduced in the case where P = M +1 by applying the expansion (see for instance PCT/EP2005/011586) ,

0154(v*,H*,Λ)^+Δ£ lie:

0155where vis a unit vector with components v_p such that Dv=O, and AE is the prediction loss energy,

0156<img file="WO2007140809A1_D0020.tif" />

0157The computation of (15) is then advantageously replaced by the application of (16) in (14), leading to

0158IA.r-1d|-Δ£ Σv Pbn.P :i8)

0159Subsequently, a preferred specialization to prediction of three channels from two channels will be discussed. The case where M = 2and P = 3 is used in MPEG Surround. The signals are a combined left X₁=I, a combined right X₂ = rand a (scaled) combined center/lfe channel X₃=C. The downmix matrix is

0160<img file="WO2007140809A1_D0021.tif" />

0161and the prediction matrix is constructed from two transmitted real parameters c_uc₂ , according to <img file="WO2007140809A1_D0022.tif" />

0162Under the assumption that the original channels are uncorreiated the prediction matrix realizing the minimal prediction error is given by

0163LC+LR -LC

01641

0165'model -RC RC+LR (21)

0166LC+RC+LR RC LC

0167Equating C_model=C leads to the (unnormalized) energy distribution taught by the present invention

0168(22) <img file="WO2007140809A1_D0023.tif" />

0169where α =(l-c,)/3, β = (\-c₂)/3 , σ = a +β, and p = aβ . This holds in the viable range defined by

0170a>0,β>0,σ<l , (23)

0171in which case the prediction error can be found in the same scaling from

0172Δ£ = 3p(l-σ). (24)

0173Since P= 3 = 2+\ =M+\ , the method outlined by (16)-(18) is applicable. The unit vector is [v,,v₂,v₃] =[l,l,-l]/>/3 and with the definitions <img file="WO2007140809A1_D0024.tif" />

0174and

0175E_n° = β{\ -σ)\k_πt ₊a^ -σ)\b_n$ ₊p\b_n$ ₍ ( 2 β ;

0176the compensation gain for each ear « = 1,2 as computed in a preferred embodiment of the gain calculator 302 can be expressed by

0177&amp;, <img file="WO2007140809A1_D0025.tif" />

0178Here ε>0 is a small number whose purpose is to stabilize the formula near the edge of the viable parameter range and g_max is an upper limit on the applied compensation gain. The gains of (27) are different for the left and right ears, « = 1,2. A variant of the method is to use a common gain g_ϋ= g_λ= g , where

0179<img file="WO2007140809A1_D0026.tif" />

0180The inventive correction gain factor can be brought into coexistence with a straight-forward multichannel gain compensation available without any HRTF related issues. In MPEG Surround, compensation for the prediction loss is already applied in the decoder by multiplying the upmix matrix C by a factor Mp where 0&lt;p≤\ is a part of the transmitted spatial parameters. In that case the gains of (27) and (28) have to be replaced by the products pg_n and pg respectively. Such compensation is applied for the binaural decoding studied in Figs. 5 and 6. It is the reason why the prior art decoding of Fig. 5 has boosted parts of the spectrum in comparison to the reference. For the subbands corresponding to those frequency regions, the inventive gain compensation effectively replaces the transmitted parameter gain factor 1/pwith a smaller value derived from formula (28).

0181In addition, since the case wherep = l corresponds to a successful prediction, a more conservative variant of the gain compensation taught by the present invention will disable the binaural gain compensation forp=l.

0182Furthermore, the present invention is used together with a residual signal. In MPEG Surround, an additional prediction residual signal z₃ can be transmitted which makes it possible to reproduce the original P=3 signals ^more faithfully. In this case the gain compensation is to be replaced by a binaural residual signal addition which will now be outlined. The predictive upmix enhanced by a residual is formed according to

0183x_p(k)=∑c_pmz_m(k)+w_p-_Zi(k)_f (29) m=\

0184where [W_pW₂₅W₃J = [I₅I₅-I]ZS . Substituting x_p for x_p in ( 5 ) yields the corresponding combined filtering,

0185Λ⁽*^{) •}*„)⁽*⁾ , ^{( 30 )}<img file="WO2007140809A1_D0027.tif" /> where the combined filters h_nmare defined by (7) for m = 1,2, and the combined filters for the residual addition are defined by

0186<img file="WO2007140809A1_D0028.tif" />

0187The overall structure of this mode of decoding is therefore also described by Fig. 2 by setting P = M = 3, and by modifying the combiner 203 to perform only the linear combination defined by (7) and (31).

0188Fig. 13 illustrates in a modified representation the result of the linear combiner 301 in Fig. 3. The result of the combiner are four HRTF-based filters hn, h_i2, h₂i and h₂₂ • As will be clearer from the description of Fig. 16a and Fig. 17, these filters correspond to filters indicated by

018915, 16, 17 ,18 in Fig. 16a.

0190Fig. 16a shows a head of a listener having a left ear or a left binaural point and having a right ear or a right binaural point. When Fig. 16a would only correspond to a stereo scenario, then filters 15, 16, 17, 18 would be typical head related transfer functions which can be individually measured or obtained via the Internet or in corresponding textbooks for different positions between a listener and the left channel speaker and the right channel speaker.

0191However, since the present invention is directed to a multi-channel binaural decoder, filters illustrated by 15,

019216, 17, 18 are not pure HRTF filters, but are HRTF-based filters, which not only reflect HRTF properties but which also depend on the spatial parameters and, particularly, as discussed in connection with Fig. 2, depend on the spatial parameter set 1 and the spatial parameter set 2. Fig. 14 shows the basis for the HRTF-based filters used in Fig. 16a. Particularly, a situation is illustrated where a listener is positioned in a sweet spot between five speakers in a five channel speaker setup which can be found, for example, in typical surround home or cinema entertainment systems. For each channel, there exist two HRTFs which can be converted to channel impulse responses of a filter having the HRTF as the transfer function. Particularly as it is known in the art, an HRTF-based filter accounts for the sound propagation within the head of a person so that, for example, HRTFl in Fig. 14 accounts for the situation that a sound emitted from speaker L_s meets the right ear after having passed around the head of the listener. Contrary thereto, the sound emitted from the left surround speaker L₃ meets the left ear almost directly and is only partly affected by the position of the ear at the head and also the shape of the ear etc. Thus, it becomes clear that the HRTFs 1 and 2 are different from each other.

0193The same is true for the HRTFs 3 and 4 for the left channel, since the relations of both ears to the left channel L are different. This also applies for all other HRTFs, although as becomes clear from Fig. 14, the HRTFs 5 and 6 for the center channel will be almost identical or even completely identical to each other, unless the individual listeners asymmetry is accommodated by the HRTF data.

0194As stated above, these HRTFs have been determined for model heads and can be downloaded for any specific "average head", and loudspeaker setup.

0195Now, as becomes clear at 171 and 172 in Fig. 17, a combination takes place to combine the left channel and the left surround channel to obtain two HRTF-based filters for the left side indicated by I/ in Fig. 15. The same procedure is performed for the right side as illustrated by R' in Fig. 15 which results in HRTF 13 and HRTF 14. To this end, reference is also made to item 173 and item 174 in Fig. 17. However, it is to be noted here that, for combining respective HRTFs in items 171, 172, 173 and 174, inter channel level difference parameters reflecting the energy distribution between the L channel and the Ls channel of the original setup or between the R channel and the Rs channel of the original multi-channel setup are accounted for. Particularly, these parameters define a weighting factor when HRTFs are linearly combined.

0196As outlined before, a phase factor can also be applied when combining HRTFs, which phase factor is defined by time delays or unwrapped phase differences between the to be combined HRTFs. However, this phase factor does not depend on the transmitted parameters.

0197Thus, HRTFs 11, 12, 13 and 14 are not true HRTFs filters but are HRTF-based filters, since these filters not only depend from the HRTFs, which are independent from the transmitted signal. Instead, HRTFs 11, 12, 13 and 14 are also dependent on the transmitted signal due to the fact that the channel level difference parameters cldχ and cld_r are used for calculating these HRTFs 11, 12, 13 and 14.

0198Now, the Fig. 15 situation is obtained, which still has three channels rather than two transmitted channels as included in a preferred down-mix signal. Therefore, a combination of the six HRTFs 11, 12, 5, 6, 13, 14 into four HRTFs 15, 16, 17, 18 as illustrated in Fig. 16a has to be done.

0199To this end, HRTFs 11, 5, 13 are combined using a left upmix rule, which becomes clear from the upmix matrix in Fig. 16b. Particularly the left upmix rule as shown in Fig 16b and as indicated in block 175 includes parameters mn, In₂I and πi3χ. This left upmix rule is in the matrix equation of Fig. 16 only for being multiplied by the left channel. Therefore, these three parameters are called the left upmix rule.

0200As outlined in block 176, the same HRTFs 11, 5, 13 are combined, but now using the right upmix rule, i.e., in the Fig. 16b embodiment, the parameters mi2, m₂2 and m₃₂, which all are used for being multiplied by the right channel Ro in Fig. 16b.

0201Thus, HRTF 15 and HRTF 17 are generated. Analogously HRTF 12, HRTF 6 and HRTF 14 of Fig. 15 are combined using the upmix left parameters mn, m₂i and m₃i to obtain HRTF 16. A corresponding combination is performed using HRTF 12, HRTF, 6 HRTF 14, but now with the upmix right parameters or right upmix rule indicated by mi₂, 1TI22 and m₃₂ to obtain HRTF 18 of Fig. 16a.

0202Again, it is emphasized that, while original HRTFs in Fig. 14 did not at all depend on the transmitted signal, the new HRTF-based filters 15, 16, 17, 18 now depend on the transmitted signal, since the spatial parameters included in the multi-channel signal were used for calculating these filters 15, 16, 17 and 18.

0203To finally obtain a binaural left channel L_B and a binaural right channel R₆, the outputs of filters 15 and 17 have to be combined in an adder 130a. Analogously, the output of the filters 16 and 18 have to be combined in an adder 130b. These adders 130a, 130b reflect the superposition of two signals within the human ear.

0204Subsequently, Fig. 18 will be discussed. Fig. 18 shows a preferred embodiment of an inventive multi-channel decoder for generating a binaural signal using a downmix signal derived from an original multi-channel signal. The downmix signal is illustrated at zi and z₂ or is also indicated by "L" and "R". Furthermore, the downmix signal has parameters associated therewith, which parameters are at least a channel level difference for left and left surround or a channel level difference for right and right surround and information on the upmixing rule.

0205Naturally, when the original multi-channel signal was only a three-channel signal, cldχ or cld_r are not transmitted and the only parametric side information will be information on the upmix rule which, as outlined before, is such an upmix rule which results in an energy-error in the upmixed signal. Thus, although the waveforms of the upmixed signals when a non-binaural rendering is performed, match as close as possible the original waveforms, the energy of the upmixed channels is different from the energy of the corresponding original channels.

0206In the preferred embodiment of Fig. 18, the upmix rule information is reflected by two upmix parameters cpci, cpc2. However, any other upmix rule information could be applied and signaled via a certain number of bits. Particularly, one could signal certain upmix scenarios and upmix parameters using a predetermined table at the decoder so that only the table indices have to be transmitted from an encoder to the decoder. Alternatively, one could also use different upmixing scenarios such as an upmix from two to more than three channels. Alternatively, one could also transmit more than two predictive upmix parameters which would then require a corresponding different downmix rule which has to fit to the upmix rule as will be discussed in more detail with respect to Fig. 20.

0207Irrespective of such a preferred embodiment for the upmix rule information, any upmix rule information is sufficient as long as an upmix to generate an energy-loss affected set of upmixed channels is possible, which is waveform-matched to the corresponding set of original signals. The inventive multi-channel decoder includes a gain factor calculator 180 for calculating at least one gain factor q_lf g_r or g, for reducing or eliminating the energy-error. The gain factor calculator calculates the gain factor based on the upmix rule information and filter characteristics of HRTF-based filters corresponding to upmix channels which would be obtained, when the upmix rule would be applied. However, as outlined before, in the binaural rendering, this upmix does not take place. Nevertheless, as discussed in connection with Fig. 15 and blocks 175, 176, 177, 178 of Fig. 17, HRTF-based filters corresponding to these upmix channels are nevertheless used.

0208As discussed before, the gain factor calculator 180 can calculate different gain factors q_x and g_r as outlined in equation (27), when, instead of n, 1 or r is inserted. Alternatively, the gain factor calculator could generate a single gain factor for both channels as indicated by equation (28) .

0209Importantly, the inventive gain factor calculator 180 calculates the gain factor based not only on the upmix rule, but also based on the filter characteristics of the HRTF-based filters corresponding to upmix channels. This reflects the situation that the filters themselves also depend on the transmitted signals and are also affected by an energy-error. Thus, the energy-error is not only caused by the upmix rule information such as the prediction parameters CPCi, CPC₂, but is also influenced by the filters themselves.

0210Therefore, for obtaining a well-adapted gain correction, the inventive gain factor not only depends on the prediction parameter but also depends on the filters corresponding to the upmix channels as well.

0211The gain factor and the downmix parameters as well as the HRTF-based filters are used in the filter processor 182 for filtering the downmix signal to obtain an energy-corrected binaural signal having a left binaural channel L_B and having a right binaural channel R_B-

0212In a preferred embodiment, the gain factor depends on a relation between the total energy included in the channel impulse responses of the filters corresponding to upmix channels to a difference between this total energy and an estimated upmix energy error ΔE. ΔE can preferably be calculated by combining the channel impulse responses of the filters corresponding to upmix channels and to then calculating the energy of the combined channel impulse response. Since all numbers in the relations for G_L and G_R in Fig. 18 are positive numbers, which becomes clear from the definitions for ΔE and E, it is clear that both gain factors are larger than 1. This reflects the experience illustrated in Fig. 5 that, in most times, the energy of the binaural signal is lower than the energy of the original multi-channel signal. It is also to note, that even when the multi-channel gain compensation is applied, i.e., when the factor p is used in most signals, nevertheless an energy-loss is caused.

0213Fig. 19a illustrates a preferred embodiment of the filter processor 182 of Fig. 18. Particularly, Fig. 19a illustrates the situation, when in block 182a the combined filters 15, 16, 17, and 18 of Fig. 16a without gain compensation are used and the filter output signals are added as outlined in Fig. 13. Then, the output of box 182a is input into a sealer box 182b for scaling the output using the gain factor calculated by box 180.

0214Alternatively, the filter processor can be constructed as shown in Fig. 19b. Here, HRTFs 15 to 18 are calculated as illustrated in box 182c. Thus, the calculator 182c performs the HRTF combination without any gain adjustment. Then, a filter adjuster 182d is provided, which uses the inventively calculated gain factor. The filter adjuster results in adjusted filters as shown in block 18Oe, where block 180e performs the filtering using the adjusted filter and performs the subsequent adding of the corresponding filter output as shown in Fig. 13. Thus, no post-scaling as in Fig 19a is necessary to obtain gain-corrected binaural channels L_B and R_B.

0215Generally, as has been outlined in connection with equation 16, equation 17 and equation 18, the gain calculation takes place using the estimated upmix error ΔE . This approximation is especially useful for the case where the number of upmix channels is equal to the number of downmix channels +1. Thus, in case of two downmix channels, this approximation works well for three upmix channels. Alternatively, when one would have three downmix channels, this approximation would also work well in a scenario in which there are four upmix channels.

0216However, it is to be noted that the calculation of the gain factor based on an estimation of the upmix error can also be performed for scenarios in which for example, five channels are predicted using three downmix channels. Alternatively, one could also use a prediction-based upmix from two downmix channels to four upmix channels. Regarding the estimated upmix energy-error ΔE, one can not only directly calculate this estimated error as indicated in equation (25) for the preferred case, but one could also transmit some information on the actually occurred upmix error in a bit stream. Nevertheless, even in other cases than the special case as illustrated in connection with equations (25) to (28) , one could then calculate the value E_n^B based on the HRTF-based filters for the upmix channels using prediction parameters. When equation (26) is considered, it becomes clear that this equation can also easily be applied to a 2/4 prediction upmix scheme, when the weighting factors for the energies of the HRTF-based filter impulse responses are correspondingly adapted. In view of that, it becomes clear that the general structure of equation (27), i.e., calculating the gain factor based on relation of E^B/ (E^B - ΔE^B) also applies for other scenarios.

0217Subsequently, Fig. 20 will be discussed to show a schematic implementation of a prediction-based encoder which could be used for generating the downmix signal L, R and the upmix rule information transmitted to a decoder so that the decoder can perform the gain compensation in the context of the binaural filter processor.

0218A downmixer 191 receives five original channels or, alternatively, three original channels as illustrated by (L_s and R₃) . The downmixer 191 can work based on a predetermined downmix rule. In that case, the downmix rule indication as illustrated by line 192 is not required. Naturally, the error- minimizer 193 could vary the downmix rule as well in order to minimize the error between reconstructed channels at the output of an upmixer 194 with respect to the corresponding original input channels.

0219Thus, the error-minimizer 193 can vary the downmix rule 192 or the upmixer rule 196 so that the reconstructed channels have a minimum prediction loss ΔE. This optimization problem is solved by any of the well-known algorithms within the error-minimizer 193, which preferably operates in a subband-wise way to minimize the difference between the reconstruction channels and the input channels.

0220As stated before, the input channels can be original channels L, L₃, R, R₃, C. Alternatively the input channels can only be three channels L, R, C, wherein, in this context, the input channels L, R, can be derived by corresponding OTT boxes illustrated in Fig. 11. Alternatively, when the original signal only has channels L, R, C, then these channels can also be termed as "original channels". Fig. 20 furthermore illustrates that any upmix rule information can be used besides the transmission of two prediction parameters as long as a decoder is in the position to perform an upmix using this upmix rule information. Thus, the upmix rule information can also be an entry into a lookup table or any other upmix related information.

0221The present invention therefore, provides an efficient way of performing binaural decoding of multi-channel audio signals based on available downmixed signals and additional control data by means of HRTF filtering. The present invention provides a solution to the problem of spectral coloring arising from the combination of predictive upmix with binaural decoding.

0222Depending on certain implementation requirements of the inventive methods, the inventive methods can be implemented in hardware or in software. The implementation can be performed using a digital storage medium, in particular a disk, DVD or a CD having electronically readable control signals stored thereon, which cooperate with a programmable computer system such that the inventive methods are performed. Generally, the present invention is, therefore, a computer program product with a program code stored on a machine readable carrier, the program code being operative for performing the inventive methods when the computer program product runs on a computer. In other words, the inventive methods are, therefore, a computer program having a program code for performing at least one of the inventive methods when the computer program runs on a computer.

0223While the foregoing has been particularly shown and described with reference to particular embodiments thereof, it will be understood by those skilled in the art that various other changes in the form and details may be made without departing from the spirit and scope thereof. It is to be understood that various changes may be made in adapting to different embodiments without departing from the broader concepts disclosed herein and comprehended by the claims that follow.