JP2007527543A - Constrained filter coding of polyphonic signals - Google Patents

Abstract

Signals of different channels (c _{1 to} c _N ) are combined into one monaural signal (x). One of the adaptive filters of a set is preferably for each channel (c ₁ to c _N), respectively of the filter adaptation unit (30: 1~30: N) is derived in. When the adaptive filter is applied to the monaural signal (x), the adaptive filter reconstructs the signal for each channel (c ₁ -c _N ) under perceptual constraints. Perceptual constraints are gain constraints and / or shape constraints. The gain constraint allows the relative energy between the channels (c ₁ -c _N ) to be maintained, while the shape constraint allows further stability by avoiding the need for spectral nulls. The parameters to be transmitted are the monaural signal (x) in coded form and the parameters (p ₁ -p _N ) of the adaptive filter that are also preferably coded. The receiver reconstructs the signals on the different channels by applying an adaptive filter and possibly some additional post-processing.

Description

  The present invention relates to encoding audio signals, and more particularly to encoding multi-channel audio signals.

  There is a high market demand for transmitting or storing audio signals at a low bit rate while maintaining high audio quality. Specifically, when there are restrictions on transmission resources and storage devices, low bit rate operation is an essential cost factor. This is a common recognition in streaming and messaging applications in mobile communication systems such as GSM, UMTS, CDMA, for example.

  At present, there are no standardized codecs that can be used in mobile communication systems that provide high stereo audio quality at economically interesting bit rates. What is possible with the current codec is mono transmission of audio signals. Stereo transmission is also available to some extent. However, due to bit rate restrictions, the current situation is that very limited stereo representations are unavoidable.

  The simplest method of stereo encoding or multi-channel encoding of an audio signal is to separately encode signals of different channels as individual independent signals. Another basic method used in stereo FM radio transmission and coexisting with a conventional monaural radio receiver is to transmit two-channel sum and difference signals.

  Current audio codecs such as MPEG-1 / 2 Layer III and MPEG-2 / 4 AAC use so-called joint stereo coding. According to this method, the signals of different channels are processed jointly rather than separately and individually. The two most commonly used joint stereo coding methods are “Mid / Side” (M / S) stereo coding and intensity stereo coding. Known as intensity stereo coding, these are generally applied to the subbands of the stereo signal or multichannel signal to be encoded.

  M / S stereo coding is described in stereo FM radio in the sense that it encodes and transmits the sum and difference signals of the channel sub-bands, thereby exploiting the redundancy between the channel and sub-bands. The procedure is the same. The structure and operation of an encoder based on M / S stereo coding is described, for example, in US Pat. No. 5,285,498 by J. D. Johnston.

  On the other hand, intensity stereo can take advantage of stereo independence. This transmits the combined strength of the channels (in different subbands) along with some position information indicating how the strength is distributed in the channel. Intensity stereo outputs only the spectral amplitude information of the channel. Phase information is not transmitted. For this reason and because time-channel information (more specifically, channel-to-channel time difference) exhibits a major psychoacoustic relevance especially at low frequencies, intensity stereo is only at frequencies above 2 kHz, for example. Can be used. Intensity stereo coding is described, for example, in European Patent No. 0497413 by R. Veldhuis et al.

  A recently developed stereo coding method is, for example, the name “Binaural cue coding applied to stereo and multi channel audio compression” by C. Faller et al. (Conference Literature). This method is a parametric multi-channel audio coding method. The basic principle is that on the encoding side, the input signals from N channels c1, c2,..., CN are combined into a single mono signal m. The mono signal is audio encoded using any conventional mono audio codec. In parallel, parameters are derived from the channel signal that describes the multi-channel image. The parameters are encoded and sent to the decoder along with the audio bit stream. The decoder first decodes the monaural signal m 'and then regenerates the channel signals c1', c2 ', ..., cN' based on the parameter description of the multichannel image.

  The principle of the binaural cue coding (BCC) method transmits an encoded mono signal and so-called BCC parameters. The BCC parameters have coded inter-channel level difference and inter-channel time difference information for the subbands of the original multi-channel input signal. The decoder regenerates the difference channel signal by applying the subband level and phase adjustment of the mono signal based on the BCC parameters. An advantage over M / S or intensity stereo etc. is that stereo information with time channel information is transmitted at a much lower bit rate.

  The problem with the above described prior art multi-channel coding method is that it requires a high bit rate in order to provide good quality. For example, when a low bit rate of about several kbps is applied, intensity stereo is adversely affected because it does not provide time channel information. This information cannot provide a stereo image at such low frequencies, for example, since low frequencies below 2 kHz are perceptually important.

  BCC can regenerate a multichannel image even at a low frequency, for example, at a low bit rate of 3 kbps. The reason is that information between time channels is also transmitted. However, this technique requires a computationally time-frequency transform for each of the channels in both the encoder and the decoder. Furthermore, BCC optimizes the mapping in a purely mathematical manner. However, the characteristic artifact inherent in the encoding method does not disappear.

  The technique described in US Pat. No. 5,434,948 by C. E. Holt et al. Uses a technique similar to BCC for encoding mono signals and side information. In this case, the sub information consists of a prediction filter and optionally a residual signal. In the prediction filter, evaluation is performed by a Least-Mean-Square algorithm, and when applied to a monaural signal, a multi-channel audio signal can be predicted. This technique can achieve very low bit rate coding for multi-channel sound sources, but with a loss of quality, as discussed further below.

  A technique similar to the above filtering technique is described in WO 03/090206 by Breebaart and Groenendaal. However, this approach uses a fixed filter that is applied to the monaural signal and combined with the monaural signal before filtering by matrix operations. The matrix operation depends on the reception correlation parameter and the reception level parameter. The purpose of such signal synthesis is to recover the correlation and level difference between the original two channels. Due to the inherent fixed filter operation, signal synthesis has very limited possibilities for signal regeneration and does not adapt to the characteristics of the signal. The approach can now be viewed as an extension of the intensity stereo coding discussed above in which the time component is communicated to the decoder. Furthermore, only the level and correlation parameters allow a certain degree of adaptation through matrix operations. This operation consists of mere rotation and scaling of the static filtered signal, thus limiting the polyphonic regeneration capability. Another drawback of this approach is that it is not based on fidelity criteria such as a signal-to-noise ratio that limits scalability to transparent quality.

  Finally, for completeness, the techniques used for 3D audio are described. This technique combines a right channel signal and a left channel signal by filtering the sound source signal with so-called head-related filters. However, this technique requires that different sound source signals be separated and is therefore generally not applicable to stereo or multi-channel coding.

US Pat. No. 5,285,498 European Patent No. 0497413 C. Faller et al., “Binaural cue coding applied to stereo and multi-channel audio compression”, 112th AES Conference, May 2002, Munich, Germany US Pat. No. 5,434,948 WO03 / 090206 European Patent 0965123B1

  Prediction filters are known to be optimal from a Least-Mean-Square perspective, but do not always fully restore the audibility characteristics of the original multi-channel signal. For example, in the case of stereo coding, stereo image instability can occur, and sound jumps randomly between left and right. In addition, spectral nulls can cause instability and can be a filter that strays the frequency response at these frequencies. This can cause the filter to perform unnecessary amplification in certain areas, which can result in very nasty noise, particularly if the signal is low or high pass filtered.

  It is an object of the present invention to provide a multi-channel coding method and apparatus that improves the perceived quality of an audio signal. Another object of the present invention is to provide the above method and apparatus at a low bit rate.

  The above objective is accomplished by a method and device according to the appended claims. In general, on the encoder side, signals from different channels are combined into one main signal. A set of adaptive filters is derived, preferably one for each channel. When the filter is applied to the main signal, the filter reconstructs the signal for each channel under perceptual constraints. Perceptual constraints are gain constraints and / or shape constraints. Gain constraints allow for the preservation of relative energy between channels, while shape constraints allow for stereo image stabilization, for example, by avoiding unnecessary filtering of spectral nulls. The transmitted parameters are the main signal in encoded form and the parameters of the adaptive filter that are also preferably encoded. The receiver reconstructs the signals on the different channels by applying an adaptive filter and possibly some additional processing.

  An advantage of the present invention is that abnormal noise when an audio signal is decoded is reduced. Also, the required transmission bit rate is maintained at a very low level.

  FIG. 1 is a diagram illustrating an exemplary system 1 according to a preferred embodiment of the present invention. The transmitter 10 includes an antenna 12 that includes associated hardware and software so that the radio signal 5 can be transmitted to the receiver 20. The transmitter 10 comprises, among other things, a multi-channel encoder 14, which converts several input channel 16 signals into output signals suitable for wireless transmission. An example of a suitable multi-channel encoder 14 will be described in detail later. The input channel 16 signal may be provided by an audio signal storage device 18, such as a digital representation data file for recording, a magnetic tape, or an audio vinyl disk recording. The signal on the input channel 16 may be provided “live” from a set of microphones 19, for example. If the audio signal is not already in digital form, it is digitized before entering multichannel encoder 14.

  On the receiver 20 side, an antenna 22 with associated hardware and software handles the actual reception of the radio signal 5 representing polyphonic audio signals. Here, normal functions such as error correction are implemented. A decoder 24 decodes the received radio signal 5, thereby converting the conveyed audio data into a number of output channel 26 signals. The output signal can be provided to the speaker 29 for immediate output, or can be stored in any type of audio signal storage device 28, for example.

  The system 1 can be, for example, a system for providing a conference call system, voice service or other audio application. In some systems, such as teleconference systems, communication must be a dual transmission system, while the distribution of musical sounds from service providers to subscribers is basically a unidirectional transmission system. It's okay. Transmission of the signal from the transmitter 10 to the receiver 20 may be performed by any other means, such as by different types of electromagnetic waves, cables, or fibers, or combinations thereof.

  FIG. 2a shows one embodiment of a multi-channel encoder 14 according to the present invention. Several channel signals c1, c2,..., CN are received at inputs 16: 1 to 16: N, respectively.

The channel signal is connected to the linear combination unit 34. In this embodiment, all channel signals are summed together to form a monaural signal x. However, any predetermined linear combination of one or more of the channel signals may be used instead, including pure channel signals. But pure sums make mathematical operations the easiest. The monaural signal x is applied as input signal 42 to the channel filter section 130. Furthermore, the monaural signal x is supplied to a monaural signal encoder 38 and encoded there to provide an encoding parameter p _x representing the monaural signal x. The monaural signal encoder operates according to any suitable monaural signal encoding technique. Many such techniques can utilize known techniques. The details of the encoding technique are not important as the enablement requirement of the present invention, and the description is omitted.

  The channel signal is also connected to the channel filter section 130. In this embodiment, each channel signal is connected to a respective filter adaptation unit 30: 1 to 30: N. The filter adaptation unit performs the reconstruction of the respective channel signal when applied to the monaural signal x. The coefficients of the filter adaptation units 30: 1 to 30: N are according to the invention optimized under perceptual constraints. However, the optimization coefficients of the filter adaptation units 30: 1 to 30: N can also be obtained at least partly in the joint optimization of two or more channel signals.

The output of the channel filter section 130 comprises N sets of filter parameters p _{1 to} p _N. These filter parameters p ₁ -p _N are usually encoded separately or together as appropriate for transmission. The filter parameters p _{1 to} p _N and the monaural signal x are sufficient to allow reconstruction of all channel signals. The encoding filter parameters p _{1 to} p _N and the encoding parameter p _x representing the monaural signal x are multiplexed with the output signal 52 in the multiplexer 40 in this embodiment so as to facilitate transmission.

FIG. 2b is a diagram illustrating one embodiment of a multi-channel decoder 24 according to the present invention. The decoder 24 of FIG. 2b is suitable for decoding the multi-channel signal encoded by the encoder of FIG. 2a. An input signal 54 is received and provided to a demultiplexer 56, which demultiplexes the input signal into several sets of encoding parameters p _x and encoding filter parameters p ₁ -p _N representing the mono signal x. To divide.

A coding parameter p _x representing the monaural signal x is provided to the monaural signal decoder 64, where the coding parameter p _x representing the monaural signal x is any arbitrary associated with the coding technique used in FIG. 2a. According to a suitable decoding technique, it is used to generate a composite mono signal x ". Many such techniques are available for known techniques. Details of the encoding technique are as an operational requirement of the present invention. The description is omitted because it is not important. The decoded mono signal x ″ is provided to the channel filter section 160.

The decoded filter parameters are also provided to the channel filter section 160 where they are decoded and used to determine the channel filters 60: 1-60: N. Each channel filter 60: 1-60: N so determined is applied to the decoded mono signal x ", thereby reconstructing each channel signal c" _1- c " _N , Provided at outputs 26: 1 to 26: N.

  In most embodiments of the present disclosure, a mono signal is used as the main signal for regenerating the channel signal in encoding or decoding. However, in a general approach, any predetermined linear combination selected from among the channel signals can be used as the main signal. The optimal choice for a given linear combination depends on the actual application field and implementation. A single channel signal may constitute such a predetermined linear combination.

  Another embodiment of a multi-channel encoder 14 according to the present invention is shown in FIG. 3a. Similar parts are denoted by similar reference numerals, and only the differences are discussed below.

The linear combiner 34 provides a predetermined linear combination of the channel signals to the monaural signal encoder 38 as before. However, in this embodiment, the signal associated with the monaural signal x is instead a decoded version x "of the encoding parameter p _x representing the monaural signal x. Such a configuration is referred to as a closed loop approach and is described below. We expect to compensate to some extent for inaccuracies in monaural signal encoding, as further described in.

The linear combination 34 of the present embodiment also combines the channel signals in _N−1 predetermined linear combinations c ^* ₁ −c _{* N−1} , which is the actual input signal to the channel filter section 130. Acts as The predetermined linear combinations c ^* ₁ −c _{* N−1} of _N−1 should be linearly independent of each other. The linear combination c ^* _1- c _{* N-1} does not necessarily have contributions from all channel signals. The term “linear combination” should be used in this context as also providing a special case where the component factor can be set to zero. In fact, in the most simple setup, linear combinations c _{^* 1} -c ^* _N-1 may be the same in the channel signal c ₁ -c _N-1. By using the monaural signal x "decoded at the decoder side, the original channel signal can be recovered.

  The modified channel signal is also connected to the channel filter section 130 in this embodiment, and N-1 sets of filter coefficients corresponding to the modified channel signal are inferred. The coefficients of the filter adaptation units 30: 1 to 30: N are according to the invention optimized under perceptual constraints.

The output of the channel filter section 130 includes N-1 sets of filter parameters p ^* _1- p ^* _N-1 . These filter parameters p ^* ₁ ~p ^* _N-1 is normally to be optimal for transmission are encoded separately or jointly. Encoding parameters p _x representing encoded filter parameters p ^* ₁ -p ^* _N-1 and the mono signal x is, in this embodiment, is sent separately.

FIG. 3b shows another embodiment of a multi-channel decoder 24 according to the present invention. The decoder of FIG. 3b is suitable for decoding the multi-channel signal encoded by the encoder of FIG. 3a. An encoding parameter p _x representing the monaural signal x and a set of encoding filter parameters p ^* _{1 to} p ^* _N−1 are received. The encoding parameter p _x representing the monaural signal x is used in the monaural signal decoder 64 to generate the decoded monaural signal x ″ as in the previous embodiment. Filter parameters p ^* _{1 to} p ^* _{N −1} is also provided to the channel filter section 160 to obtain _N−1 decoded modified channel signals c ^* _{1 to} c ^* _N−1 , then linear combination 74 is applied to the modified channel Used to provide channel signals c " _{1 to} c" _N reconstructed from signals c ^* _{1 to} c ^* _N-1 and decoded monaural signal x ".

  In order to understand the important relevance of perceptual constraints, an example of prior art filter coding is described in more detail, basically referring to US Pat. No. 5,434,948. This multi-channel coding allows a low bit rate when transmission of residual signals is omitted. In order to derive a channel reconstruction filter, an error minimization procedure based on the concept of LMS or weighted LMS ensures that the filter output signal ^ c (n) is optimally matched to the target signal c (n). Calculate

Several error measures may be used to calculate the filter. Mean square error or weighted mean square error is well known and requires a small amount of calculation. According to the LMS method, the filter h ^c _uc is effective for one frame of data, and “uc” indicates “unconstrained”, and the square error between the target signal and the filter output, that is, the difference r _uc (n) = c (n) _-chosen to minimize the square of ^ c _uc (n). Here, n is an index of data frame samples. This error is expressed as follows.

As a result, the following linear system is obtained for the filter coefficient vector h ^c _uc .

は、モノラル信号x(n)の対称共分散行列である。 However,

Is the symmetric covariance matrix of the monaural signal x (n).

In the above equation, rxc is a vector of cross-correlation between signals x (n) and c (n).

  However, as mentioned above, the perceptual characteristics may not be completely determined by pure mathematical minimization.

  One very important perceptual characteristic of a multichannel signal is the energy and specifically the relative level between the multichannel audio signals. In the case of stereo coding according to prior art methods, a troublesome stereo image instability in which the sound source jumps periodically from left to right can result. In addition, direct control over left and right predictions is not achieved because only one filter is required in stereo coding. According to the invention, it is therefore advantageous that gain constraints are used during the optimization procedure. Note that in this context, one filter per channel is essentially required (see FIGS. 2a and 2b above).

  In certain situations, the predicted channel may not have frequency content above or below a certain frequency. This occurs, for example, when the channel is high pass filtered or obtained from a band splitting procedure. Spectral nulls can cause instability and result in a filter response that results in unwanted amplification and low frequency noise. Thus, according to the present invention, shape constraints are advantageously used during the optimization procedure.

これから、出力信号^c1(n)を導出する。ファクタγ_c1およびγ_c2は、チャネル信号がどのように組み合わされるかを決定する。１つの可能性は、γ_c1をファクタ２γに設定し、γ_c2を２（１−γ）に設定するものである。この場合、モノラル信号は、チャネルの重み付け和である。具体的には、適切なセッティングがγ＝０．５であり、この場合、両チャネルとも、等しく重み付けされる。他の適切なセッティングは、γ_c1＝−γ_c2とすることが可能であり、この場合、モノラル信号は、チャネル信号の差である。 FIG. 4 shows the basic concept of the encoder side constraint minimization procedure according to the invention in an embodiment with two channels (in the case of stereo) and a linear filter 31. A filter 31 responsive to the constraint of channel c1 with filter coefficient h _c1 is derived in optimization unit 32 by a constraint error minimization procedure. The filter h _c1 takes as input the combined channel signal, which in this embodiment is a linear combination of the two channel signals c1 and c2, ie the monaural signal x (n),

From this, the output signal ^ c1 (n) is derived. The factors γ _c1 and γ _c2 determine how the channel signals are combined. One possibility is to set γ _c1 to a factor 2γ and γ _c2 to 2 (1-γ). In this case, the monaural signal is a weighted sum of the channels. Specifically, an appropriate setting is γ = 0.5, where both channels are equally weighted. Another suitable setting may be γ _c1 = −γ _c2 where the mono signal is the difference of the channel signals.

  The weighted combination of individual channel signals to form a mono signal can generally even be a combination of filtered versions of each channel signal. Such an approach is called pre-filtering. This may be useful if the approach is implemented in the excitation domain or generally in the weighted signal domain. For example, the channel can be pre-filtered with an LPC (Linear Predictive Coding) residual filter of the mono signal.

  In the following, the mono and left and right channels are generally assumed to be some pre-filtered versions of the actual mono, left and right channels. When recovering the channel, a pre-filtering step with a mono LPC synthesis filter is necessary to return to the signal domain.

In the following, the cases γ _c1 = 1/2 and γ _c2 = 1/2 will be discussed in more detail.

If h _c1 is a FIR (finite impulse response) filter, ^ c1 (n) is a linear combination of delayed versions of the signal x (n).

The index set is I = [i _min Ki _max ]. The filter parameter p ₁ may be the additional data needed to determine the filter with the filter coefficient h _c1 .

  For example, when applying the encoding method disclosed in US Pat. No. 5,434,948, the difference signal of the two channel signals is regenerated by a filter. In FIG. 5, the right and left signals are indicated by curves 301 and 302, respectively. It is assumed that the display is not ideal and gives a slightly larger difference than the target difference over the entire frame. As a result, a regenerated right signal 303 slightly lower than the original right signal is obtained on the decoder side, and a regenerated left signal 304 slightly higher than the original left signal is obtained. Such abnormal noise is a decrease in the volume of the right channel and an increase in the volume of the left channel. If such an abnormal sound changes with time, the sound will swing back and forth between the right and left channels. Gain constraints can improve such situations.

There are several ways to implement gain constraints. One possible approach is to have a hard constraint, ie an exact energy match between the original channel and the evaluation channel, or the output channel has a defined energy E _c1 that is not necessarily equal to the original channel signal energy. Such a loose gain constraint is imposed.

  The problem of constraint minimization can be easily solved by the Lagrangian method, that is, the Lagrangian function.

The optimal solution gives the filter h _c1 , which is proportional to the next unconstrained filter.

  The proportional factor is:

Therefore, the gain constraint filter is
h _c1 ^gc = g _c1 h _c1 ^uc
It becomes.

If this coding principle is used in a limited frequency band, the channel signal may look like the curve 305 in FIG. The intensity does not exist below the frequency f ₁ or above the frequency f ₂ . However, pure mathematical optimization results in curve 306, which also manifests some limited power below and above frequencies f ₁ and f ₂ , respectively. Such artifacts are perceived.

  In order to impose a certain spectral shape on the filter, a set of linear constraints must be imposed on the filter. These constraints should generally be less than the number of filter coefficients.

  For example, if it is desired to set a spectral null constraint at 0 kHz, a suitable constraint is:

  In general, the shape of the constraint can be formulated by a matrix and a vector as follows:

  From the constraint least-squares theory, the optimal filter that satisfies these constraints is:

  This constraint is particularly useful when it is empirically known that the channel has no frequency content in certain frequency regions.

  Gain and shape constraints can also be combined. In such a case, it is preferred that the shape constraint is first applied and then the gain constraint is added as a factor according to the following.

  Since the filter depends on the unconstrained filter and the latter is obedient, the relationship is as follows by c1 (n) + c2 (n) = 2x (n).

  Here, δ represents the same filter (identity filter). If the constraints on the two channels are the same, useful properties can be derived for the shape constraint filter.

Therefore, it becomes as follows.

  This equation is useful for reducing the bit rate when coding the channel filter, because the channel filter is related by the amount available at the decoder side. It is.

  The relationship between shape constraint filters also applies to the rational calculation of filters. FIG. 7 shows that one of the two channels c 1 and c 2 is regenerated by applying the monaural signal x to the unconstrained filter 131. The result of the unconstrained filter is modified according to the shape constraints in the shape constraint section 132. From the c1 channel shape constraint filter, a channel c2 shape constraint filter may also be calculated and provided to separate the gain constraint section 133 for each channel.

A more detailed block scheme of another embodiment that uses side signals to apply shape constraints is shown in FIG. The two channel signals c1 and c2 are combined into the monaural signal x and the sub signal s in the adding means 55 and 57 of the linear combination unit 34. The channel filter section 130 comprises an unconstrained parameter filter 131, which is applied to the monaural signal x to regenerate the evaluation of the side signal ^ s. In the unconstrained optimization unit 33, the filter coefficients are adapted to give a minimum difference between s and ^ s. The filter h _s ^uc thus obtained is provided to the shape constraint section 132 basically according to the further discussion above. A side signal shape constraint filter h _s ^sc is generated. From the relationship (1) between the channel filters in the stereo application field, the shape constraint filter of each channel signal is calculated based on the shape constraint filter h _s ^sc of the sub signal. These filters, not coefficients, are provided in the respective constraint sections 133: 1, 133: 2. The gain factor for each channel signal is calculated and the two filters are provided to the parameter encoding section 66 where the parameters of the two filters are encoded jointly.

After computing the constrained channel filters h _c1 and h _c2 , they are quantized and encoded in an appropriate representation to be transmitted to the receiver. Typically, the filter coefficients are quantized using a scalar or vector quantizer and the quantizer index is transmitted. The quantizer can also perform predictions that are very beneficial for bit rate reduction, especially in this scenario.

By utilizing the complementarity of the filter, it is possible to further reduce the bit rate because only one of the filters h _c1 or h _c2 or their linear combination is quantized and transmitted, On the other hand, gains g _c1 and g _c2 are jointly vector quantized and transmitted separately. Such transmission can be carried out at a bit rate as low as 1 kbps, for example.

The receiver first decodes the transmitted monaural signal and channel filter. The receiver then regenerates a different channel signal by filtering the mono signal with the respective channel filter. In the case of stereo, the complementary property is used and the coefficients are preferably _recombined to produce filters h _c1 and h _c2 .

  Certain post-processing steps that further improve the quality of the reconstructed multi-channel signal can follow the regeneration of the different channel signals.

  It may be beneficial to smooth the gain of the shape constrained filters or the linear combination of these filters before calculating the gain constrained channel filters.

  For example, for stereo, the equivalent sub-signal filter is as follows (as used in FIG. 8):

  In order to reduce abnormal noise, the gain difference of this filter between successive frames is smoothed to obtain the following filter:

  The channel filter is then modified according to the following.

  Although this type of modification does not maintain the shape constraint, it can be readily recognized that the shape constraint is still maintained for the sub-signal filter, which is sufficient for stereo coding.

The gain constraint for the filter assumes the previously calculated channel energy, ie E _c1 , E _c2 . It is important to avoid unnecessary amplification by controlling the gain of filters such as g _c1 and g _c2 and limiting the gain. Depending on the characteristics of the different channel signals, it can take place that the channels are cross-correlated for the entire frequency range or in a certain frequency band. This leads to some erasure when a monaural signal is formed. In this case, individual channel information is lost, at least in part and in some frequency bands, so it is beneficial to limit the channel gain when the channel gain is greater than some amount, such as 0 dB. One way to implement this gain limitation is to calculate a certain gain factor.

The above equation is the ratio of the effective mono channel energy to the mono channel energy when the two channels are not correlated. When this factor is less than 0 dB, it has signal cancellation. In this case, g _F quantifies how severe this elimination is. The gain information can then be calculated as follows:

  The same restriction applies to the gain of other channels.

  Not only the channel filter parameters need to be encoded and transmitted, but also a mono signal. When deriving channel filter coefficients, there are two different principle approaches that take into account monaural signal speech coding.

  In the open loop scheme, the filter is derived based on the original mono signal. This is the case for example in FIG. 2a, where the signal 42 is the original monaural signal x. However, the decoder uses the quantized monaural signal as the input of the channel filter.

  In the closed-loop scheme, the filter calculation is based on a monaural signal that has been coded and thus already quantized. This is the case, for example, in FIG. 3a, where the signal 44 is an encoded mono signal x ". This approach only allows the channel filter design to match each channel signal in an optimally possible manner. This technique is also aimed at reducing the coding error resulting from the monaural signal coding.

The principles described so far are applicable to full spectrum, ie full band signals. However, it is equally or more beneficial that this is applicable to the signal sub-bands. FIG. 9 illustrates the principle of sub-band processing. Several channels c _{1 to} c _N are respectively divided in the K sub-bands SB1, SB2 and SBk. Each sub-band channel signal is output to a respective multi-channel encoder 80: 1-80: K, where the channel signal is encoded. One or several of the multi-channel encoders 80: 1 to 80: K may be multi-channel encoders according to the present invention. A bitstream combiner 82 combines the encoded signal with the common encoded signal 53, which is transmitted.

  An advantage of the sub-band processing described above is that multi-channel coding of different sub-channels can be performed individually and optimized, for example with respect to allocation bit rate, processing frame size, and sampling rate.

  One special type of sub-band processing does not perform multi-channel coding for very low frequencies, eg, below 200 Hz. This means that a simple monaural signal is transmitted in this very low frequency band. This principle takes advantage of the fact that human stereo perception is less sensitive at very low frequencies. Known from prior art and subwoofers.

  In other embodiments of sub-band processing, band splitting is performed using a time-frequency transform, such as a short-time Fourier transform (STFT) that allows the signal to be decomposed into signal frequency components, for example. In this case, the filtering is simply a multiplication of the individual spectral coefficients of the monaural signal and the complex coefficients.

  The parametric multi-channel coding method according to the present invention typically includes fixed frame processing of signal samples. That is, parameters describing the multi-channel image are derived and transmitted at a rate corresponding to the encoded frame length such as 20 ms. However, the parameters can be obtained from signal frames that are much longer than the coding frame length. A suitable option is to set the length of such an analysis frame to a value longer than the encoded frame length. This indicates that the parameter calculation is performed in the overlap analysis frame.

  This is illustrated in FIG. The encoder analysis frame 83 is slightly longer than the encoding frame 84, as shown at the top of the figure. As a result of such an overlap analysis frame, the parameters transition smoothly. This is essential for providing a stable multi-channel audio signal image. The same is performed at the decoder side, as shown in the middle of the figure. Therefore, in consideration of this, it is indispensable for the decoder to add a duplication with an overlap 86 by adding a window to the synthesized frame 85 as shown at the bottom of the figure. This makes it possible to smoothly transition between filters associated with each frame.

  In the encoder, the development of smoothing filter parameters can be enhanced. For example, low pass or intermediate filtration can be applied to the filter parameters.

  Prior art mono speech codecs, as well as speech codecs, perform so-called noise shaping of coding noise. The purpose of this operation is to move the coding noise to a frequency where the signal has a high spectral density and thus makes the noise less audible. Noise shaping is usually performed adaptively, i.e. in response to an audio signal. This generally means that the noise shaping performed on a mono signal is different from that required for various channel signals. As a result, despite proper noise shaping in the mono speech codec, the subsequent channel filtering according to the present invention can be applied to the audible coding of the reconstructed multi-channel signal when compared to the audible coding noise of the mono signal. May increase.

  To alleviate this problem, a signal adaptive post filter can be applied to the reconstructed channel signal in the receiver post-processing step. In this case, any prior art post-filtering technique can be used, which emphasizes spectral peaks or deepens spectral valleys, thereby reducing audible noise. An example of such a technique is E.I. It is described in European Patent 0965123B1 by Ekudden et al. Another simple method is the so-called pitch postfilter and formant postfilter known from speech coding.

  In FIG. 11, the main steps of one embodiment of the encoding method according to the invention are shown in a flowchart. The procedure begins at step 200. In step 220, the main signal, preferably a monaural signal inferred from the multi-channel signal, is encoded. In step 222, the filter coefficients are optimized to give the best possible representation of the channel signal when applied to the main signal. Optimization is performed under perceptual constraints. The optimal coefficients are then encoded at step 224. The procedure ends at step 299.

  The above-described embodiments should be understood as some examples of the invention. Those skilled in the art will be able to make various modifications, combinations and changes to the above-described embodiments without departing from the scope of the present invention. In particular, different partial solutions of different embodiments can be combined in other configurations where technically possible. In any case, the scope of the present invention is defined by the appended claims.

1 is a block diagram of a system for transmitting a multi-channel signal. FIG. 4 is a block diagram of an embodiment of an encoder in a transmitter according to the present invention. FIG. 4 is a block diagram of an embodiment of a decoder in a receiver according to the present invention. FIG. 6 is a block diagram of another embodiment of an encoder in a transmitter according to the present invention. FIG. 6 is a block diagram of another embodiment of a decoder in a receiver according to the present invention. FIG. 4 is a block diagram of an embodiment of a filter adaptation unit according to the present invention. FIG. 6 illustrates the effect of insufficient regeneration of sub-signals in a prior art system. It is a figure which shows the influence of the spectrum null in the system of a prior art. FIG. 6 is a block diagram illustrating combinability in a channel filter section according to the present invention. FIG. 3 is a block diagram of an embodiment of an encoder that uses partial combinatorial coding of stereo signals. FIG. 6 is a block diagram illustrating the use of division in frequency sub-bands. It is a compound figure which shows the overlap analysis of encoding and decoding. 4 is a flowchart of basic steps of an embodiment of an encoding method according to the present invention;

Claims (18)

A method for encoding a multi-channel signal (c ₁ -c _N ) having at least a first channel and a second channel, comprising:
Generating a coding parameter (p _x ) representing a main signal (x) that is a first predetermined linear combination of signals of the multichannel signals (c _{1 to} c _N );
Deriving optimal parameters (p _{1 to} p _N ) of the first adaptive filter (31; 131, 132, 133: 1-2);
Encoding the optimal parameters (p _{1 to} p _N ),
Further comprising deriving optimal parameters (p _{1 to} p _N ) of at least the second adaptive filter (31; 131, 132, 133: 1-2);
When the first adaptive filter (31; 131, 132, 133: 1-2) is applied to the first predetermined linear combination (x), the first adaptive filter (31; 131, 132, 133: 1- 2) is derived to give the minimum difference between the signal of the first channel (c _{1 to} c _N ) and the filter output signal,
The minimum difference is determined according to the first criterion,
When the second adaptive filter is applied to the first predetermined linear combination (x), the second adaptive filter calculates a minimum difference between the signal of the second channel (c _{1 to} c _N ) and the filter output signal. Derived to give,
The minimum difference is determined according to the second criterion,
Each derivation step of the first and second adaptive filters (31; 131,132,133: 1-2) is performed under at least one perceptual constraint selected from the group of gain constraints and shape constraints A method characterized by.   The method of claim 1, wherein at least one of the first criterion and the second criterion is a Least-Mean-Square (LMS) criterion.   Method according to claim 1 or 2, characterized in that the perceptual constraint is at least a gain constraint and is intended to give the total energy of the filter output signal equal to the total energy of the signal of the first channel. .   4. The gain constraint of claim 3, wherein the gain constraint is an absolute constraint and requires that the total energy of the adaptive filter output signal is equal to the total energy of the corresponding channel signal. Method.   The gain constraint is a soft constraint, supporting that the adaptive filter provides the total energy of the adaptive filter output signal close to the total energy of the signal of the corresponding channel. 3. The method according to 3. 4. The method of claim 3, wherein the gain constraint is imposed as a gain factor (g _{c1 to} g _cN ) multiplied by a filter derived without having a gain constraint. The method of claim 6, wherein the gain constraint filter h _c ^gc is given by:

Where h _c ^uc is an adaptive filter derived without gain constraints, E _c is the specified energy of the adaptive filter output signal, and ^ c ^uc (n) is the main signal x (n) without gain constraints Is an adaptive filter output.   The method according to claim 1, wherein the perceptual constraint is at least a shape constraint and imposes a predetermined spectral shape on the adaptive filter.   The method of claim 8, wherein the shape constraint imposes null content in a predetermined frequency range. It said step of encoding the optimal parameters (p ₁ ~p _N) is in any one of claims 1 to 9, characterized in that encode together said optimal parameters of said first and second filters The method described. Said step of deriving parameters comprises:
Generating ^{_{a; (c * 1 ~c * N}} -1 s), the multi-channel signal a second predetermined linear combination of (c ₁ ~c _N) of the signal
When the third filter is applied to the first predetermined linear combination under the shape constraint, the third filter is configured to provide a minimum difference between the second predetermined linear combination and the filter output signal. Deriving parameters of
11. The method according to claim 1, comprising calculating the optimal parameters of the first and second filters as a function of the optimal parameters of the third filter. Step A method according to any one of claims 1 to 11, wherein said be performed based on the encoding parameters (p _x) representing the main signal (x) to derive the parameters.   12. A method according to any of the preceding claims, wherein the step of deriving parameters is performed based on the first predetermined linear combination (x).   The multi-channel signal has three or more channels, and the main signal is based on all first predetermined linear combinations (x) of the three or more channels, and the signal of each channel is 14. A method according to any of claims 1 to 13, characterized in that it is represented by a separate adaptive filter optimized under perceptual constraints. A method for decoding a polyphonic signal having a coding parameter (p _x ) representing a main signal and a coding optimal parameter of a first adaptive filter (60: 1 to 60: N) comprising:
Decoding the encoding parameter (p _x ) representing the main signal;
Applying the first adaptive filter (60: 1 to 60: N) to the decoded main signal (x ") to generate a signal of a first channel (c" _{1 to} c " _N ); Have
The encoding parameter (p _x ) further represents an encoding optimal parameter of the second adaptive filter (60: 1 to 60: N),
Applying the second adaptive filter (60: 1 to 60: N) to the decoded main signal (x ") to generate a signal of a second channel (c" _{1 to} c " _N ); In addition,
The method wherein the first and second adaptive filters (30: 1 to 30: N) are optimized under at least one perceptual constraint selected from a group of gain constraints and shape constraints. As a predetermined linear combination of the decoded main signal (x ") and the first channel (c" _{1 to} c " _N ), the second channel (c" _{1 to} c " _N ) The method of claim 15, comprising generating. Input means (16: 1 to 16: N) for inputting a multi-channel signal (c _{1 to} c _N ) comprising at least a first channel and a second channel;
An encoding parameter (p _x ) connected to the input means (16: 1 to 16: N) and representing a main signal (x) that is a first predetermined linear combination of the signals of the multichannel signals (c _{1 to} c _N ). ) (38)
Means for deriving optimal parameters of the first adaptive filter (31; 131, 132, 133: 1-2);
Means (66) for encoding said optimal parameters;
Output means (52) and
Means (31; 131, 132, 133: 1-2) for deriving optimal parameters of the second adaptive filter;
When the first adaptive filter is applied to the first predetermined linear combination (x), the first adaptive filter calculates a minimum difference between the signal of the first channel (c _{1 to} c _N ) and the filter output signal. Give,
The minimum difference is determined according to the first criterion,
When the second adaptive filter is applied to the first predetermined linear combination (x), the second adaptive filter is the minimum of the signal of the second channel (c _{1 to} c _N ) and the filter output signal. Give the difference
The minimum difference is determined according to a second criterion,
Each means (31; 131, 132, 133: 1-2) for deriving optimal parameters of the first and second adaptive filters is subject to at least one perceptual constraint selected from a group of gain constraints and shape constraints. An encoding device (14), which is configured to derive the optimum parameter. Input means (54) for inputting a coding parameter representing the main signal (x) and a coding optimum parameter of the first adaptive filter;
Means (64) for decoding the encoding parameter (p _x ) representing the main signal (x);
Means (60: 1 to 60: N) for generating a first channel signal by applying the first adaptive filter to the decoded main signal (x ");
The encoding parameter (p _x ) further represents an encoding optimal parameter of the second adaptive filter (60: 1 to 60: N);
Means for generating a second channel signal by applying the second adaptive filter to the decoded main signal (x ");
The decoding device (24), wherein the first and second adaptive filters are constrained under at least one perceptual constraint selected from a group of a gain constraint and a shape constraint.

Images