Advanced processing based on a complex-exponential-modulated filterbank and adaptive time signalling methods

TECHNICAL FIELD

0001The present invention relates to audio source coding systems but the same methods could also be applied in many other technical fields. Different techniques that are useful for audio coding systems using parametric representations of stereo properties are introduced.

BACKGROUND OF THE INVENTION AND PRIOR ART

0002The present invention relates to parametric coding of the stereo image of an audio signal. Typical parameters used for describing stereo image properties are inter-channel intensity difference (IID), inter-channel time difference (ITD), and inter-channel coherence (IC). In order to re-construct the stereo image based on these parameters, a method is required that can re-construct the correct level of correlation between the two channels, according to the IC parameter. This is accomplished by a de-correlation method.

0003There are a couple of methods available for creation of decorrelated signals. Ideally, a linear time invariant (LTI) function with all-pass frequency response is desired. One obvious method for achieving this is by using a constant delay. However, using a delay, or any other LTI all-pass functions, will result in non-all-pass response after adding the non-processed signal. In the case of a delay, the result will be a typical comb-filter. The comb-filter often gives an undesirable "metallic" sound that, even if the stereo widening effect can be efficient, reduces much naturalness of the original.

0004Frequency domain methods for generating a de-correlated signal by adding a random sequence to the IID values along the frequency axis, where different sequences are used for the different audio channels, are also known from prior art. One problem with frequency domain decorrelation by the random sequence modifications is the introduction of pre-echoes. Subjective tests have shown that for non-stationary signals, pre-echoes are by far more annoying than post-echoes, which is also well supported by established psycho acoustical principles. This problem could be reduced by dynamically adapting transform sizes to the signal characteristics in terms of transient content. However, switching transform sizes is always a hard (i.e., binary) decision that affects the full signal bandwidth and that can be difficult to accomplish in a robust manner.

0005United States patent application publication <patcit id="pcit0001" dnum="US20030219130A1"><text>US 2003/0219130 A1</text></patcit> discloses a coherence-based audio coding and synthesis. In particular, an auditory scene is synthesized from a mono audio signal by modifying, for each critical band, an auditory scene parameter such as an inter-aural level difference (ILD) and/or an inter-aural time difference (ITD) for each subband within the critical band, where the modification is based on an average estimated coherence for the critical band. The coherence-based modification produces auditory scenes having object widths, which more accurately match the widths of the objects in the original input auditory scene. Stereo parameters are the well-known BCC parameters, wherein BCC stands for binaural cue coding. When generating two different decorrelated output channels, frequency coefficients as obtained by a discrete Fourier transform are grouped together in a single critical band. Based on the inter-channel coherence measure, weighting factors are multiplied by a pseudo-random sequence which is preferably chosen such that the variance is approximately constant for all critical bands, and the average is "0" within each critical band. The same sequence is applied to the spectral coefficients of each different frame.

SUMMARY OF THE INVENTION

0006It is the object of the present invention to provide a decoding concept for parametrically encoded multi-channel signals or an encoding concept for generating such signals which result in a good audio quality and a good coding efficiency.

0007This object is achieved by an apparatus for generating a decorrelation signal in accordance with claim 1, a multi-channel decoder in accordance with claim 13, a method of generating a decorrelation signal in accordance with claim 20, a method of multi-channel decoding in accordance with claim 21, an apparatus for encoding a stereo signal in accordance with claim 22 or a method of encoding a stereo signal in accordance with claim 26 or a computer program in accordance with claim 27.

0008The present invention is based on the finding that, on the decoding side, a good decorrelation signal for generating a first and a second channel of a multi-channel signal based on the input mono signal is obtained, when a reverberation filter is used, which introduces an integer or preferably a fractional delay into the input signal. Importantly, this reverberation filter is not applied to the whole input signal. Instead, several reverberation filters are applied to several subbands of the original input signal, i.e., the mono signal so that the reverberation filtering using the reverberation filters is not applied in a time domain or in the frequency domain, i.e., in the domain which is reached, when a Fourier transform is applied. Inventively, the reverberation filtering using reverberation filters for the subbands is individually performed in the subband domain.

0009A subband signal includes a sequence of at least two subband samples, the sequence of the subband samples representing a bandwidth of the subband signal, which is smaller than the bandwidth of the input signal. Naturally, the frequency bandwidth of a subband signal is higher than a frequency bandwidth attributed to a frequency coefficient obtained by Fourier transform. The subband signals are preferably generated by means of a filterbank having for example 32 or 64 filterbank channels, while an FFT would have, for the same example, 1.024 or 2.048 frequency coefficients, i.e., frequency channels.

0010The subband signals can be subband signals obtained by subband-filtering a block of samples of the input signal. Alternatively, the subband filterbank can also be applied continuously without a block wise processing. For the present invention, however, block wise processing is preferred.

0011Since the reverberation filtering is not applied to the whole signal, but is applied subband-wise, a "metallic" sound caused by comb-filtering is avoided.

0012In cases, in which a sample period between two subsequent subband samples of the subband is too large for a good sound impression at the decoder end, it is preferred to use fractional delays in a reverberation filter such as a delay between 0.1 and 0.9 and preferably 0.2 to 0.8 of the sampling period of the subband signal. It is noted that in case of critical sampling, and when 64 subband signals are generated using a filterbank having 64 filterbank channels, the sampling period in a subband signal is 64 times larger than the sampling period of the original input signal.

0013It is to be noted here that the delays are an integral part of the filtering process used in the reverberation device. The output signal constitutes of a multitude of delayed versions of the input signal. It is preferred to delay signals by fractions of the subband sampling period, in order to achieve a good reverberation device in the subband domain.

0014In preferred embodiments of the present invention, the delay, and preferably the fractional delay introduced by each reverberation filter in each subband is equal for all subbands. Nevertheless, the filter coefficients are different for each subbands. It is preferred to use IIR filters. Depending on the actual situation, fractional delay and the filter coefficients for the different filters can be determined empirically using listening tests.

0015The subbands filtered by the set of reverberation filters constitute a decorrelation signal which is to be mixed with the original input signal, i.e., the mono signal to obtain a decoded left channel and decoded right channel. This mixing of a decorrelation signal with the original signal is performed based on an inter-channel coherence parameter transmitted together with the parametrically encoded signal. To obtain different left and right channels, i.e., different first and second channels, mixing of the decorrelation signal with a mono signal to obtain the first output channel is different from mixing the decorrelation signal with the mono signal to obtain the second output channel.

0016To obtain higher efficiency on the encoding side, multi-channel encoding is performed using an adaptive determination of the stereo parameter set. To this end, an encoder includes, in addition to a means for calculating the mono signal and in addition to a means for generating a stereo parameter set, a means for determining a validity of stereo parameter sets for subsequent portions of the left an right channels. Preferably, the means for determining is operative to activate the means for generating, when it is determined that the stereo parameter set is not valid anymore so that a second stereo parameter set is calculated for portions of the left and right channels starting at a second time border. This second time border is also determined by the means for determining a validity.

0017The encoded output signal then includes the mono signal, a first stereo parameter set and a first time border associated with the first parameter set and the second stereo parameter set and the second time border associated with the second stereo parameter set. On the decoding side, the decoder will use a valid stereo parameter set until a new time border is reached. When this new time border is reached, the decoding operations are performed using the new stereo parameter set.

0018Compared to prior art methods, which did a block wise processing and, therefore, a block wise determination of stereo parameter sets, the inventive adaptive determination of stereo parameter sets for different encoder-side determined time borders provides a high coding efficiency on the one hand end and a high coding quality on the other hand. This is due to the fact that for relatively stationary signals, the same stereo parameter set can be used for many blocks of the samples of the mono signal without introducing audible errors. On the other hand, when non-stationary signals are concerned, the inventive adaptive stereo parameter determination provides an improved time resolution so that each signal portion has its optimum stereo parameter set.

0019The present invention provides a solution to the prior art problems by using a reverberation unit as a de-correlator implemented with fractional delay lines in a filterbank, and using adaptive level adjustment of the de-correlated reverberated signal.

0020Subsequently, several aspects of the present invention are outlined.

0021One aspect of the invention is a method for delaying a signal by: filtering a real-valued time domain signal through the analysis part of complex filterbank; modifying the complex-valued subband signals obtained from the filtering; and filtering the modified complex-valued subband signals through the synthesis part of the filterbank; and taking the real part of the complex-valued time domain output signal, where the output signal is the sum of the signals obtained from the synthesis filtering.

0022Another aspect of the invention is a method for modifying the complex valued subband signals by filtering each complex-valued subband signal with a complex valued finite impulse response filter where the finite impulse response filter for subband number <i>n</i> is given by a discrete time Fourier transform of the form <maths id="math0001" num=""><math display="inline"><msub><mi>H</mi><mi>n</mi></msub><mfenced><mi>ω</mi></mfenced><mo>=</mo><mrow><mo>{</mo><mtable columnalign="left"><mtr><mtd><mi>exp</mi><mfenced separators=""><mo>-</mo><mi mathvariant="italic">iπ</mi><mo>⁢</mo><mfenced separators=""><mi>n</mi><mo>+</mo><mn>1</mn><mo>/</mo><mn>2</mn></mfenced><mo>⁢</mo><mi>τ</mi></mfenced><mo>⁢</mo><msub><mi>G</mi><mi>τ</mi></msub><mfenced><mi>ω</mi></mfenced><mo>,</mo></mtd><mtd><mi>for</mi><mspace width="1em" /><mi>n</mi><mspace width="1em" /><mi>even</mi><mo>;</mo></mtd></mtr><mtr><mtd><mi>exp</mi><mfenced separators=""><mo>-</mo><mi mathvariant="italic">iπ</mi><mo>⁢</mo><mfenced separators=""><mi>n</mi><mo>+</mo><mn>1</mn><mo>/</mo><mn>2</mn></mfenced><mo>⁢</mo><mi>τ</mi></mfenced><mo>⁢</mo><msub><mi>G</mi><mi>τ</mi></msub><mo>⁢</mo><mfenced separators=""><mi>ω</mi><mo>+</mo><mi>π</mi></mfenced><mo>,</mo></mtd><mtd><mi>for</mi><mspace width="1em" /><mi>n</mi><mspace width="1em" /><mi>odd</mi><mn>.</mn></mtd></mtr></mtable></mrow><mo>,</mo></math><img file="EP2265040A2_D0001.tif" /></maths> where the parameter τ=<i>T</i>/<i>L</i>, and where the synthesis filter bank has <i>L</i> subbands and the desired delay is <i>T</i> measured in output signal sample units.

0023Another aspect of the invention is a method for modifying the complex valued subband signals by filtering where the filter <i>G</i>_τ(ω)approximately satisfies <i>V</i>_τ(ω)<i>G</i>_τ(ω)+<i>V</i>_τ(ω+π)<i>G</i>_τ(ω+π)=1, where <i>V</i>_τ(ω) is the discrete time Fourier transform of the sequence <maths id="math0002" num=""><math display="inline"><msub><mi>v</mi><mi>τ</mi></msub><mfenced><mi>k</mi></mfenced><mo>=</mo><mi>A</mi><mo>⁢</mo><msup><mi>i</mi><mi>k</mi></msup><mstyle displaystyle="true"><munder><mo>∑</mo><mi>l</mi></munder><mo>⁢</mo><mi>p</mi></mstyle><mfenced><mi>l</mi></mfenced><mo>⁢</mo><mi>p</mi><mo>⁢</mo><mfenced separators=""><mi>l</mi><mo>-</mo><mi>T</mi><mo>-</mo><mi mathvariant="italic">Lk</mi></mfenced><mo>,</mo></math><img file="EP2265040A2_D0002.tif" /></maths> and <i>p</i>(<i>l</i>)is the prototype filter of said complex filterbank and <i>A</i> is an appropriate real normalization factor.

0024Another aspect of the invention is a method for modifying the complex valued subband signals by filtering where the filter <i>G</i>_τ(ω)satisfies <i>G</i>_τ(-ω)=<i>G</i>_τ(ω+π)^* such that even indexed impulse response samples are real valued and odd indexed impulse response samples are purely imaginary valued.

0025Another aspect of the invention is a method for coding of stereo properties of an input signal, by at an encoder, calculate time grid parameters describing the location in time for each stereo parameter set, where the number of stereo parameter sets are arbitrary, and at a decoder, applying parametric stereo synthesis according to that time grid.

0026Another aspect of the invention is a method for coding of stereo properties of an input signal, where the time localisation for the first stereo parameter set is, in the case of where a time cue for the stereo parameter set coincides with the beginning of a frame, signalled explicitly instead of signalling the time pointer.

0027Another aspect of the invention is a method for generation of stereo decorrelation for parametric stereo reconstruction, by at a decoder, applying an artificial reverberation process to synthesise the side signal.

0028Another aspect of the invention is a method for generation of stereo decorrelation for parametric stereo reconstruction by, at the decoder, the reverberation process is made within a complex modulated filterbank using phase delay adjustment in each filter bank channel.

0029Another aspect of the invention is a method for generation of stereo decorrelation for parametric stereo reconstruction by, at the decoder, the reverberation process utilises a detector designed for finding signals where the reverberation tail could be unwanted and let the reverberation tail be attenuated or removed.

BRIEF DESCRIPTION OF THE DRAWINGS

0030The 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: <dl id="dl0001"><dt>Fig. 1</dt><dd>illustrates a block diagram of the inventive appara- tus;</dd><dt>Fig. 2</dt><dd>illustrates a block diagram of the means for generat- ing a de-correlated signal;</dd><dt>Fig. 3</dt><dd>illustrates the analysis of a single channel and the synthesis of the stereo channel pair based on the re- constructed stereo subband-signals according to the present invention;</dd><dt>Fig. 4</dt><dd>illustrates a block diagram of the division of the parametric stereo parameters sets into time segments, based on the signal characteristic; and</dd><dt>Fig. 5</dt><dd>illustrates an example of the division of the para- metric stereo parameters sets into time segments, based on the signal characteristic.</dd></dl>

DESCRIPTION OF PREFERRED EMBODIMENTS

0031The below-described embodiments are merely illustrative for the principles of the present invention for parametric stereo coding. 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.

0032Delaying a signal by a fraction of a sample can be achieved by several prior art interpolation methods. However, special cases arises when the original signal is available as oversampled complex valued samples. Performing fractional delay in the qmf bank by only applying phase delay by a factor for, each qmf channel corresponding to a constant time delay, results in severe artefacts.

0033This can efficiently be avoided by using a compensation filter according to a novel approach allowing high quality approximations to arbitrary delays in any complex-exponential-modulated filterbank. A detailed description follows below.

A continuous time model

0034For ease of computations a complex exponential modulated <i>L-</i>band filterbank will be modelled here by a continuous time windowed transform using the synthesis waveforms <maths id="math0003" num="(1)"><math display="block"><msub><mi>u</mi><mrow><mi>n</mi><mo>,</mo><mi>k</mi></mrow></msub><mfenced><mi>t</mi></mfenced><mo>=</mo><mi>v</mi><mo>⁢</mo><mfenced separators=""><mi>t</mi><mo>-</mo><mi>k</mi></mfenced><mo>⁢</mo><mi>exp</mi><mfenced open="[" close="]" separators=""><mi mathvariant="italic">iπ</mi><mo>⁢</mo><mfenced separators=""><mi>n</mi><mo>+</mo><mn>1</mn><mo>/</mo><mn>2</mn></mfenced><mo>⁢</mo><mfenced separators=""><mi>t</mi><mo>-</mo><mi>k</mi><mo>+</mo><mi>θ</mi></mfenced></mfenced><mo>,</mo></math><img file="EP2265040A2_D0003.tif" /></maths> where n,kare integers with <i>n</i>≥0 and θ is a fixed phase term. Results for discrete-time signals are obtain by suitable sampling of the <i>t</i>-variable with spacing 1/<i>L</i>. It is assumed that the real valued window <i>v</i>(<i>t</i>) is chosen such that for real valued signals <i>x</i>(<i>t</i>) it holds to very high precision that <maths id="math0004" num="(2)"><math display="block"><mi>x</mi><mfenced><mi>t</mi></mfenced><mo>=</mo><mn>2</mn><mspace width="1em" /><mi>Re</mi><mfenced open="{" close="}" separators=""><mstyle displaystyle="true"><munderover><mo>∑</mo><mrow><mi>n</mi><mo>=</mo><mn>0</mn></mrow><mi>∞</mi></munderover></mstyle><mstyle displaystyle="true"><munderover><mo>∑</mo><mrow><mi>k</mi><mo>=</mo><mo>-</mo><mi>∞</mi></mrow><mi>∞</mi></munderover></mstyle><msub><mi>c</mi><mi>n</mi></msub><mfenced><mi>k</mi></mfenced><mo>⁢</mo><msub><mi>u</mi><mrow><mi>n</mi><mo>,</mo><mi>k</mi></mrow></msub><mfenced><mi>t</mi></mfenced></mfenced></math><img file="EP2265040A2_D0004.tif" /></maths><maths id="math0005" num="(3)"><math display="block"><mi>if</mi><mspace width="1em" /><msub><mi>c</mi><mi>n</mi></msub><mfenced><mi>k</mi></mfenced><mo>=</mo><munderover><mo>∫</mo><mrow><mo>-</mo><mi>∞</mi></mrow><mi>∞</mi></munderover><mi>x</mi><mfenced><mi>t</mi></mfenced><mo>⁢</mo><msubsup><mi>u</mi><mrow><mi>n</mi><mo>,</mo><mi>k</mi></mrow><mo>*</mo></msubsup><mfenced><mi>t</mi></mfenced><mo>⁢</mo><mi mathvariant="italic">dt</mi><mo>,</mo></math><img file="EP2265040A2_D0005.tif" /></maths> where * denotes complex conjugation. It is also assumed that <i>v</i>(<i>t</i>) is essentially band limited to the frequency interval [-π,π]. Consider the modification of each frequency band <i>n</i> by filtering the discrete time analysis samples <i>c_n</i>(<i>k</i>) with a filter with impulse response <i>h_n</i>(<i>k</i>), <maths id="math0006" num="(4)"><math display="block"><msub><mi>d</mi><mi>n</mi></msub><mfenced><mi>k</mi></mfenced><mo>=</mo><mstyle displaystyle="false"><mstyle displaystyle="true"><munder><mo>∑</mo><mi>l</mi></munder></mstyle><msub><mi>h</mi><mi>n</mi></msub></mstyle><mfenced><mi>l</mi></mfenced><mo>⁢</mo><msub><mi>c</mi><mi>n</mi></msub><mo>⁢</mo><mfenced separators=""><mi>k</mi><mo>-</mo><mi>l</mi></mfenced><mn>.</mn></math><img file="EP2265040A2_D0006.tif" /></maths> Then the modified synthesis <maths id="math0007" num="(5)"><math display="block"><mi>y</mi><mfenced><mi>t</mi></mfenced><mo>=</mo><mn>2</mn><mspace width="1em" /><mi>Re</mi><mfenced open="{" close="}" separators=""><mstyle displaystyle="true"><munderover><mo>∑</mo><mrow><mi>n</mi><mo>=</mo><mn>0</mn></mrow><mi>∞</mi></munderover></mstyle><mstyle displaystyle="true"><munderover><mo>∑</mo><mrow><mi>k</mi><mo>=</mo><mo>-</mo><mi>∞</mi></mrow><mi>∞</mi></munderover></mstyle><msub><mi>d</mi><mi>n</mi></msub><mfenced><mi>k</mi></mfenced><mo>⁢</mo><msub><mi>u</mi><mrow><mi>n</mi><mo>,</mo><mi>k</mi></mrow></msub><mfenced><mi>t</mi></mfenced></mfenced></math><img file="EP2265040A2_D0007.tif" /></maths> can be computed in the frequency domain to be <maths id="math0008" num="(6)"><math display="block"><mover><mi>y</mi><mo>^</mo></mover><mfenced><mi>ω</mi></mfenced><mo>=</mo><mi>H</mi><mfenced><mi>ω</mi></mfenced><mo>⁢</mo><mover><mi>x</mi><mo>^</mo></mover><mfenced><mi>ω</mi></mfenced><mo>,</mo></math><img file="EP2265040A2_D0008.tif" /></maths> where <i>f̂</i>(ω) denotes Fourier transforms of <i>f</i>(<i>t</i>) and <maths id="math0009" num="(7)"><math display="block"><mi>H</mi><mfenced><mi>ω</mi></mfenced><mo>=</mo><mstyle displaystyle="true"><munderover><mo>∑</mo><mrow><mi>n</mi><mo>=</mo><mo>-</mo><mi>∞</mi></mrow><mi>∞</mi></munderover></mstyle><msub><mi>H</mi><mi>n</mi></msub><mfenced><mi>ω</mi></mfenced><mo>⁢</mo><msup><mfenced open="|" close="|" separators=""><mover><mi>v</mi><mo>^</mo></mover><mo>⁢</mo><mfenced separators=""><mi>ω</mi><mo>-</mo><mi>π</mi><mo>⁢</mo><mfenced separators=""><mi>n</mi><mo>+</mo><mn>1</mn><mo>/</mo><mn>2</mn></mfenced></mfenced></mfenced><mn>2</mn></msup><mn>.</mn></math><img file="EP2265040A2_D0009.tif" /></maths>

0035Here, <i>H_n</i>(ω)=∑<i>_k^h_n</i>(<i>k</i>)exp(-<i>ik</i>ω) is the discrete time Fourier transform of the filter applied in frequency band <i>n</i> for <i>n</i>≥0 and <maths id="math0010" num="(8)"><math display="block"><msub><mi>H</mi><mi>n</mi></msub><mfenced><mi>ω</mi></mfenced><mo>=</mo><msub><mi>H</mi><mrow><mo>-</mo><mn>1</mn><mo>-</mo><mi>n</mi></mrow></msub><mo>⁢</mo><msup><mfenced separators=""><mo>-</mo><mi>ω</mi></mfenced><mo>*</mo></msup><mspace width="1em" /><mi>for</mi><mspace width="1em" /><mi>n</mi><mo>&lt;</mo><mn>0.</mn></math><img file="EP2265040A2_D0010.tif" /></maths>

0036Observe here that the special case <i>H_n</i>(ω)=1 leads to <i>H</i>(ω)=1 in (7) due to the special design of the window v(<i>t</i>). Another case of interest is <i>H_n</i>(ω)=exp(-<i>i</i>ω)which gives <i>H</i>(ω)=exp(-<i>i</i>ω), so that <i>y</i>(<i>t</i>) = <i>x</i>(<i>t</i>-1).

The proposed solution

0037In order to achieve a delay of size τ, such that <i>y</i>(<i>t</i>)=<i>x</i>(<i>t</i>-τ), the problem is to design filters <i>H_n</i>(ω) for <i>n</i>≥0 such that <maths id="math0011" num="(9)"><math display="block"><mi mathvariant="italic">H</mi><mfenced><mi mathvariant="normal">ω</mi></mfenced><mo>=</mo><mi>exp</mi><mfenced separators=""><mo>−</mo><mi mathvariant="italic">iτω</mi></mfenced><mo>,</mo></math><img file="EP2265040A2_D0011.tif" /></maths> where <i>H</i>(ω) is given by (7) and (8). The particular solution proposed here is to apply the filters <maths id="math0012" num="(10)"><math display="block"><msub><mi>H</mi><mi>n</mi></msub><mfenced><mi>ω</mi></mfenced><mo>=</mo><mrow><mo>{</mo><mtable columnalign="left"><mtr><mtd><mi>exp</mi><mfenced separators=""><mo>-</mo><mi mathvariant="italic">iπ</mi><mo>⁢</mo><mfenced separators=""><mi>n</mi><mo>+</mo><mn>1</mn><mo>/</mo><mn>2</mn></mfenced><mo>⁢</mo><mi>τ</mi></mfenced><mo>⁢</mo><msub><mi>G</mi><mi>τ</mi></msub><mfenced><mi>ω</mi></mfenced><mo>,</mo></mtd><mtd><mi>for</mi><mspace width="1em" /><mi>n</mi><mspace width="1em" /><mi>even</mi><mo>;</mo></mtd></mtr><mtr><mtd><mi>exp</mi><mfenced separators=""><mo>-</mo><mi mathvariant="italic">iπ</mi><mo>⁢</mo><mfenced separators=""><mi>n</mi><mo>+</mo><mn>1</mn><mo>/</mo><mn>2</mn></mfenced><mo>⁢</mo><mi>τ</mi></mfenced><mo>⁢</mo><msub><mi>G</mi><mi>τ</mi></msub><mo>⁢</mo><mfenced separators=""><mi>ω</mi><mo>+</mo><mi>π</mi></mfenced><mo>,</mo></mtd><mtd><mi>for</mi><mspace width="1em" /><mi>n</mi><mspace width="1em" /><mi>odd</mi><mn>.</mn></mtd></mtr></mtable></mrow></math><img file="EP2265040A2_D0012.tif" /></maths>

0038Here <i>G</i>_τ(-ω)=<i>G</i>_τ(ω+π)* implies consistency with (8) for all <i>n.</i> Insertion of (10) into the right hand side of (7) yields <maths id="math0013" num="(11)"><math display="block"><mi>H</mi><mfenced><mi>ω</mi></mfenced><mo>=</mo><mi>exp</mi><mfenced separators=""><mo>-</mo><mi mathvariant="italic">iωτ</mi></mfenced><mo>⁢</mo><mfenced open="[" close="]" separators=""><msub><mi>V</mi><mi>τ</mi></msub><mfenced><mi>ω</mi></mfenced><mo>⁢</mo><msub><mi>G</mi><mi>τ</mi></msub><mfenced><mi>ω</mi></mfenced><mo>+</mo><msub><mi>V</mi><mi>τ</mi></msub><mo>⁢</mo><mfenced separators=""><mi>ω</mi><mo>+</mo><mi>π</mi></mfenced><mo>⁢</mo><msub><mi>G</mi><mi>τ</mi></msub><mo>⁢</mo><mfenced separators=""><mi>ω</mi><mo>+</mo><mi>π</mi></mfenced></mfenced></math><img file="EP2265040A2_D0013.tif" /></maths> where <i>V</i>_τ(ω)=∑<i>_nb</i>(ω-π(2<i>n</i>+1/2)) with <i>b</i>(ω)=exp(<i>i</i>τω)|<i>v̂</i>(ω)|². Elementary computations show that <i>V</i>_τ(ω) is the discrete time Fourier transform of <maths id="math0014" num="(12)"><math display="block"><msub><mi>v</mi><mi>τ</mi></msub><mfenced><mi>k</mi></mfenced><mo>=</mo><msup><mi>i</mi><mi>k</mi></msup><munderover><mo>∫</mo><mrow><mo>-</mo><mi>∞</mi></mrow><mi>∞</mi></munderover><mi>v</mi><mfenced><mi>t</mi></mfenced><mo>⁢</mo><mi>v</mi><mo>⁢</mo><mfenced separators=""><mi>t</mi><mo>-</mo><mi>τ</mi><mo>-</mo><mi>k</mi></mfenced><mo>⁢</mo><mi mathvariant="italic">dt</mi><mn>.</mn></math><img file="EP2265040A2_D0014.tif" /></maths>

0039Very good approximations to the perfect delay can be obtained by solving the linear system <maths id="math0015" num="(13)"><math display="block"><msub><mi>V</mi><mi>τ</mi></msub><mfenced><mi>ω</mi></mfenced><mo>⁢</mo><msub><mi>G</mi><mi>τ</mi></msub><mfenced><mi>ω</mi></mfenced><mo>+</mo><msub><mi>V</mi><mi>τ</mi></msub><mo>⁢</mo><mfenced separators=""><mi>ω</mi><mo>+</mo><mi>π</mi></mfenced><mo>⁢</mo><msub><mi>G</mi><mi>τ</mi></msub><mo>⁢</mo><mfenced separators=""><mi>ω</mi><mo>+</mo><mi>π</mi></mfenced><mo>=</mo><mn>1</mn></math><img file="EP2265040A2_D0015.tif" /></maths> in the least squares sense with a FIR filter <maths id="math0016" num=""><math display="inline"><msub><mi>G</mi><mi>τ</mi></msub><mfenced><mi>ω</mi></mfenced><mo>=</mo><mstyle displaystyle="false"><munderover><mo>∑</mo><mrow><mi>k</mi><mo>=</mo><mo>-</mo><mi>N</mi></mrow><mi>M</mi></munderover><msub><mi>g</mi><mi>τ</mi></msub></mstyle><mfenced><mi>k</mi></mfenced><mo>⁢</mo><mi>exp</mi><mfenced separators=""><mo>-</mo><mi mathvariant="italic">ikω</mi></mfenced><mn>.</mn></math><img file="EP2265040A2_D0016.tif" /></maths> In terms of filter coefficients, the equation (13) can be written <maths id="math0017" num="(14)"><math display="block"><mn>2</mn><mstyle displaystyle="false"><mstyle displaystyle="true"><munder><mo>∑</mo><mi>l</mi></munder></mstyle><msub><mi>v</mi><mi>τ</mi></msub></mstyle><mfenced separators=""><mn>2</mn><mo>⁢</mo><mi>k</mi><mo>-</mo><mi>l</mi></mfenced><mo>⁢</mo><msub><mi>g</mi><mi>τ</mi></msub><mfenced><mi>l</mi></mfenced><mo>=</mo><mi>δ</mi><mfenced open="[" close="]"><mi>k</mi></mfenced><mo>,</mo></math><img file="EP2265040A2_D0017.tif" /></maths> where δ[<i>k</i>]=1 for <i>k</i> = 0 and δ[<i>k</i>]=0 for k≠0.

0040In the case of a discrete time <i>L</i>-band filter bank with prototype filter <i>p</i>(<i>k</i>), the obtained delay in sample units is <i>L</i>τ and the computation (12) is replaced by <maths id="math0018" num="(15)"><math display="block"><msub><mi>v</mi><mi>τ</mi></msub><mfenced><mi>k</mi></mfenced><mo>=</mo><msup><mi>i</mi><mi>k</mi></msup><mstyle displaystyle="false"><mstyle displaystyle="true"><munder><mo>∑</mo><mi>l</mi></munder></mstyle><mi>p</mi><mfenced><mi>l</mi></mfenced><mo>⁢</mo><mi>p</mi><mo>⁢</mo><mfenced separators=""><mi>l</mi><mo>-</mo><mi>T</mi><mo>-</mo><mi mathvariant="italic">Lk</mi></mfenced></mstyle><mo>,</mo></math><img file="EP2265040A2_D0018.tif" /></maths> where <i>T</i> is the integer closest to <i>L</i>τ. Here <i>p</i>(<i>k</i>) is extended by zeros outside its support. For a finite length prototype filter, only finitely many <i>v</i>_τ(<i>k</i>) are different from zero, and (14) is system of linear equations. The number of unknowns <i>g</i>_τ(<i>k</i>) is typically chosen to be a small number. For good QMF filter bank designs, 3-4 taps already give very good delay performance. Moreover, the dependence of the filter taps <i>g</i>_τ(<i>k</i>) on the delay parameter τ can often be modelled successfully by low order polynomials.

Signalling adaptive time grid for stereo parameters

0041Parametric stereo systems always leads to compromises in terms of limited time or frequency resolution in order to minimise conveyed data. It is however well known from psycho-acoustics that some spatial cues can be more important than others, which leads to the possibility to discard the less important cues. Hence, the time resolution does not have to be constant. Great gain in bitrate could be achieved by letting the time grid synchronise with the spatial cues. It can easily be done by sending a variable number of parameter sets for each data frame that corresponds to a time segment of fixed size. In order to synchronise the parameter sets with corresponding spatial cues, additional time grid data describing the location in time for each parameter set has to be sent. The resolution of those time pointers could be chosen to be quite low to keep the total amount of data minimised. A special case where a time cue for a parameter set coincides with the beginning of a frame could be signalled explicitly to avoid sending that time pointer.

0042<figref idref="f0004">Fig. 4</figref> illustrates an inventive apparatus for performing parameter analysis for time segments having variable and signal dependant time borders. The inventive apparatus includes means 401 for dividing the input signal into one or several time segments. The time borders that separate the time segments are provided by means 402. Means 402 uses a detector specially designed for extracting spatial cues that is relevant for deciding where to set the time borders. Means 401 outputs all the input signal divided into one or several time segments. This output is input to means 403 for separate parameter analysis for each time segment. Means 403 outputs one parameter set per time segment being analysed.

0043<figref idref="f0005">Fig. 5</figref> illustrates an example of how the time grid generator can perform for a hypothetical input signal. In this example one parameter set per data frame is used if no other time border information is present. Hence, when no other time border information is present, the inherent time borders of the data frame is used. The in <figref idref="f0005">Fig. 5</figref> depicted time borders are the output from means 402 in <figref idref="f0004">Fig. 4</figref>. The in <figref idref="f0005">Fig. 5</figref> depicted time segments are provided by means 401 in <figref idref="f0004">Fig. 4</figref>.

0044The apparatus for encoding a stereo signal to obtain a mono output signal and the stereo parameter set includes the means for calculating the mono signal by combining a left and a right channel of the stereo signals by weighted addition. Additionally, a means 403 are generating a first stereo parameter set using a portion of the left channel and a portion of the right channel, the portions starting at a first time border is connected to the means for determining the validity of the first stereo parameter set for subsequent portions of the left channel and the right channel.

0045The means for determining is collectively formed by the means 402 and 401 in <figref idref="f0001">Fig. 1</figref>.

0046Particularly, the means for determining is operative to generate a second time border and to activate the means for generating, when it is determined that this first stereo parameter set is not valid anymore so that a second stereo parameter set for portions of the left and right channels starting at the second time border is generated.

0047Not shown in <figref idref="f0004">Fig. 4</figref> are means for outputting the mono signal, the first stereo parameter set and the first time border associated with the first stereo parameter set and the second stereo parameter set and the second time border associated with the second stereo parameter set as the parametrically encoded stereo signal. The means for determining a validity of a stereo parameter set can include a transient detector, since the probability is high that, after a transient, a new stereo parameter has to be generated, since a signal has changed its shape significantly. Alternatively, the means for determining a validity can include an analysis-by-synthesis device, which is adapted for decoding the mono signal and the stereo parameter set to obtain a decoded left and a decoded right channel, to compare the decoded left channel and the decoded right channel to the left channel and to the right channel, and to activate the means for generating, when the decoded left channel and the decoded right channel are different from the left channel and the right channel by more than the predetermined threshold.

0048Data frame 1: The time segment corresponding to parameter set 1 starts at the beginning of data frame 1 since no other time border information is present in this data frame.

0049Data frame 2: Two time borders are present in this data frame. The time segment corresponding to parameter set 2 starts at the first time border in this data frame. The time segment corresponding to parameter set 3 starts at the second time border in this data frame.

0050Data frame 3: One time border is present in this data frame. The time segment corresponding to parameter set 4 starts at the time border in this data frame.

0051Data frame 4: One time border is present in this data frame. This time border coincides with the start border of the data frame 4 and does not have to be signalled since this is handled by the default case. Hence, this time border signal can be removed. The time segment corresponding to parameter set 5 starts at the beginning of data frame 4, even without signalling this time border.

Using artificial reverberation as decorrelation method for parametric stereo reconstruction

0052One vital part of making the stereo synthesis in a parametric stereo system is to decrease the coherence between the left and right channel in order to create wideness of the stereo image. It can be done by adding a filtered version of the original mono signal to the side signal, where the side and mono signal is defined by: <maths id="math0019" num=""><math display="block"><mi>mono</mi><mo>=</mo><mfenced separators=""><mi>left</mi><mo>+</mo><mi>right</mi></mfenced><mo>/</mo><mn>2</mn><mo>,</mo></math><img file="EP2265040A2_D0019.tif" /></maths> and <maths id="math0020" num=""><math display="block"><mi>side</mi><mo>=</mo><mfenced separators=""><mi>left</mi><mo>-</mo><mi>right</mi></mfenced><mo>/</mo><mn>2</mn><mo>,</mo></math><img file="EP2265040A2_D0020.tif" /></maths> respectively.

0053In order to not change the timbre too much, the filter in question should preferably be of all-pass character. One successful approach is to use similar all-pass filters used for artificial reverberation processes. Artificial reverberation algorithms usually requires high time resolution to give an impulse response that is satisfactory diffuse in time. There are great advantages in basing an artificial reverberation algorithm on a complex filter bank such as the complex qmf bank. The filterbank provides excellent possibilities to let the reverberation properties be frequency selective in terms of for example reverberation equalisation, decay time, density and timbre. However, the filter bank implementations usually exchanges time resolution for higher frequency resolution which normally makes it hard to implement a reverberation process that is smooth enough in time. To deal with this problem a novel method would be to use a fractional delay approximation by only applying phase delay by a factor for, each qmf channel corresponding to a constant time delay. This primitive fractional delay method introduces severe time smearing that fortunately is very much desired in this case. The time smearing contributes to the time diffusion which is highly desirable for reverberation algorithms and gets bigger as the phase delay approaches pi/2 or - pi/2.

0054Artificial reverberation processes are for natural reasons processes with an infinite impulse response, and offers natural exponential decays. In [<patcit id="pcit0002" dnum="SE0201372W"><text>PCT/SE02/01372</text></patcit>] it is pointed out that if a reverberation unit is used for generating a stereo signal, the reverberation decay might sometimes be unwanted after the very end of a sound. These unwanted reverb-tails can however easily be attenuated or completely removed by just altering the gain of the reverb signal. A detector designed for finding sound endings can be used for that purpose. If the reverberation unit generates artefacts at some specific signals e.g., transients, a detector for those signals can also be used for attenuating the same.

0055<figref idref="f0001">Fig. 1</figref> illustrates an inventive apparatus for the decorrelation method of signals as used in a parametric stereo system. The inventive apparatus includes means 101 for providing a plurality of subband signals. The providing means can be a complex QMF filterbank, where every signal is associated with a subband index.

0056The subband signals output by the means 101 from <figref idref="f0001">Fig 1</figref>. are input into a means 102 for providing a de-correlated signal 102, and into a means 103 and 106 for modifying the subband signal. The output from 102 is input into a means 104 and 105 for modifying the of the signal, and the output of 103, 104, 105 and 106 are input into a means for adding, 107 and 108, the subband signals.

0057In the presently described embodiment of the invention, the means for modifying 103,104, 105 and 106, the subband signals, adjusts the level of the de-correlated signal and the un-processed signal being the output of 101, by multiplying the subband signal with a gain factor, so that every sum of every pair results in a signal with the amount of de-correlated signal given by the control parameters. It should be noted that the gain factors used in the means for modifying, 103 - 106, are not limited to a positive value. It can also be a negative value.

0058The output from the means for adding subband signals 107 and 108, is input to the means for providing a time-domain signal 109 and 110. The output from 109 corresponds to the left channel of the re-constructed stereo signal, and the output from 110 corresponds to the right channel of the re-constructed stereo signal. In the here described embodiment the same de-correlator is used for both output channels, while the means for adding the de-correlated signal with the un-processed signal are separate for the two output channels. The presently described embodiment thereby ensures that the two output signals can be identical as well as completely de-correlated, dependent on the control data provided to the means for adjusting the levels of the signals, and the control data provided to the means for adding the signals.

0059In <figref idref="f0002">Fig. 2</figref> a block diagram of the means for providing a decorrelated signal is displayed. The input subband signal is input to the means for filtering a subband signal 201. In the presently described embodiment of the present invention the filtering step is a reverberation unit incorporating all-pass filtering. The filter coefficients used are given by the means for providing filter coefficients 202. The subband index of the currently processed subband signal is input to 202. In one embodiment of the present invention different filter coefficients are calculated based on the subband index provided to 202. The filtering step in 201, relies on delayed samples of the input subband signal as well as delayed samples of intermediate signals in the filtering procedure.

0060It is an essential feature of the present invention that means for providing integer subband sample delay and fractional subband sample delay are provided by 203. The output of 201 is input to a means for adjusting the level of the subband signal 204, and also to a means for estimating signal characteristics of the subband signal 205. In a preferred embodiment of the present invention the characteristics estimated is the transient behaviour of the subband signal. In this embodiment a detected transient is signalled to the means for adjusting the level of a subband signal 204, so that the level of the signal is reduced during transient passages. The output from 204 is the de-correlated signal input to 104 and 105 of <figref idref="f0001">Fig. 1</figref>.

0061In <figref idref="f0003">Fig. 3</figref> the single analysis filterbank and the two synthesis filterbanks are shown. The analysis filterbank 301, operates on the mono input signal, while the synthesis filterbanks 302 and 303 operate on the re-constructed stereo signals.

0062<figref idref="f0001">Fig. 1</figref>, therefore, shows the inventive apparatus for generating a decorrelation signal which is indicated by reference 102. As it is shown in <figref idref="f0001">Fig. 1</figref> or <figref idref="f0003">3</figref>, this apparatus includes means for providing a plurality of subband signals, wherein a subband signal includes the sequence of at least two subband samples, the sequence of the subband samples representing a bandwidth of the subband signal which is smaller than a bandwidth of the input signal. Each subband signal is input into the means 201 for filtering. Each means 201 for filtering includes a reverberation filter so that a plurality of reverberated subband signals are obtained, wherein the plurality of reverberated subband signals together represent the decorrelation signal. Preferably, as it is shown in <figref idref="f0002">Fig. 2</figref>, there can be a subband-wise postprocessing of reverberated subband signals which is performed by block 204, which is controlled by block 205.

0063Each reverberation filter is set to a certain delay, and preferably a fractional delay, and each reverberation filter has several filter coefficients, which, as it is shown in <figref idref="f0002">Fig. 2</figref>, depend on the subband index. This means that it is preferred to use the same delay for each subband but to use different sets of filter coefficients for the different subbands. This is symbolized by means 203 and 202 in <figref idref="f0002">Fig. 2</figref>, although it is to be mentioned here that delays and filter coefficients are preferably fixedly determined when shipping a decorrelation device, wherein the delays and filter coefficients may be determined empirically using listening tests etc.

0064A multi-channel decoder is shown by <figref idref="f0001">Fig. 1</figref> and includes the inventive apparatus for generating the correlation signal, which is termed 102 in <figref idref="f0001">Fig. 1</figref>. The multi-channel decoder shown in <figref idref="f0001">Fig. 1</figref> is for decoding a mono signal and an associated inter-channel coherence measure, the inter-channel coherence measure representing a coherence between a plurality of original channels, wherein the mono signal is derived from the plurality of original channels. Block 102 in <figref idref="f0001">Fig. 1</figref> constitutes a generator for generating a decorrelation signal for the mono signal. Blocks 103, 104, 105, 106 and 107 and 108 constitute a mixer for mixing the mono signal and the decorrelation signal in accordance with the first mixing mode to obtain a first decoded output signal and in accordance with the second mixing mode to obtain a second decoded output signal, wherein the mixer is operative to determine the first mixing mode and the second mixing mode based on the inter-channel coherence measure transmitted as a side information to the mono signal.

0065The mixer is preferably operative to mix in a subband domain based on separate inter-channel coherence measures for different subbands. In this case, the multi-channel decoder further comprises means 109 and 110 for converting the first and second decoded output signals from the subband domain in a time domain to obtain a first decoded output signal and a second decoded output signal in the time domain. Therefore, the inventive means 102 for generating a decorrelation signal and the inventive multi-channel decoder as shown in <figref idref="f0001">Fig. 1</figref> operate in the subband domain and perform, as the very last step, a subband domain to time domain conversion.

0066Depending on the actual situation, the inventive device can be implemented in hardware or in software or in a firmware including hardware constituents and software constituents. When implemented in software partially or fully, the invention also is a computer program having a computer-readable code for carrying out the inventive methods when running on a computer.