EP1926082A1 - Process for scaleable encoding of stereo signals - Google Patents

Abstract

The method involves transforming stereo signals of left and right channels (10, 20) before center/side processing from a time domain into a frequency domain. The signals are separately quantized according to matrixing for sum and difference formations for compression of data and according to the transformation in the frequency range. The quantized signals of the channels are matrixed for forming center and side signals (40, 50). The center and side signals formed from the quantized signals are utilized in a stage of loss-less encoding (400) for transmission in an encoded signal.

Description

The present invention relates to the encoding of stereo signals, and more particularly to the application of scalable encoding techniques.

Scalable encoding methods for data compression of audio signals have the advantage that the transmission rate can be adapted dynamically to the properties of the networks and terminals. A gradation of the bit rate by the coding method in small steps is particularly advantageous.

$$\mathrm{S}\mathrm{=}\mathrm{0}\mathrm{,}\mathrm{5}\left(\mathrm{R}-\mathrm{L}\right)$$

A stereo signal includes at least two channels, a left channel and a right channel. For a data-reducing coding, the similarity between the two channels is exploited. One known method of transmitting stereo signals is the mid / side method [ Michael Dickreiter, Handbuch der Tonstudiotechnik, Saur Verlag, 1997 ]. In this case, the left and right channels are combined with each other to produce a center channel and a side channel. The center channel is made up of the sum of the right and left channels, while the side channel is made up of the difference between the left and right channels. As an equation, this means: $$\mathrm{M}\mathrm{=}\mathrm{0}\mathrm{.}\mathrm{5}\left(\mathrm{R}\mathrm{+}\mathrm{L}\right)$$

$$\mathrm{S}\mathrm{=}\mathrm{0}\mathrm{.}\mathrm{5}\left(\mathrm{R}-\mathrm{L}\right)$$

$$\mathrm{L}\mathrm{=}\mathrm{M}-\mathrm{S}$$

The factor 0.5 is a common size in practice and can also be chosen differently. The recovery of the right and left channels then takes place from the relationship $$\mathrm{R}\mathrm{=}\mathrm{M}\mathrm{+}\mathrm{S}$$

$$\mathrm{L}\mathrm{=}\mathrm{M}-\mathrm{S}$$

If the left channel and the right channel are relatively similar to one another, center / side processing will result in a significant saving in the amount of bits needed for encoding, since the side channel will then have relatively less energy than the left or right channel and To code the page channel much less bits are needed. In the limit case where the left channel and the right channel are identical, the center channel will be equal to the left channel or equal to the right channel, while the side channel would be 0. The more similar the left and right channels are, the less energy the page channel will be, and the fewer bits needed to encode the page channel. If the right and left channels are less similar, then the bit efficiency will decrease accordingly for center / page encoding.

In the prior art, the coding of the stereo signals is usually carried out with methods that process the audio signals in the spectral range. First, the left and right channels of the audio signal, which are usually in the form of PCM (Pulse Code Modulation) samples, are converted from the time domain to the frequency domain. For this transformation, for example, modern coding methods use the so-called modified discrete cosine transformation (MDCT) in order to obtain a block-wise frequency representation of an audio signal. The stream of discrete-time audio samples is windowed to obtain a windowed block of audio samples, which are then transformed into a spectral representation by a transform. For each time window, one obtains a corresponding number of spectral coefficients. The transformation divides the frequency spectrum into a certain number of frequency bands (subbands) of equal width. The number of transformation points and the sampling rate determine the bandwidth of the subbands. These subbands are grouped according to hearing characteristics. At low frequencies, a few subbands fall into one group, many at high frequencies. For each group a scaling factor is determined. The quantization of the spectral coefficients then takes place relative to these scaling factors. During encoding, bits are assigned to the scaling factors and the transform coefficient according to the target bit rate. The bit allocation takes place in such a way that the resulting error can be perceived as little as possible. The scaling factors are also transmitted and are required so that the decoder is able to reconstruct the original signal from the transmitted bits.

In the case of a middle / side coding, for the signals of the left and right channels after the transformation into the frequency range MDCT is used for matrixing and subtraction. The center and side signals thus formed are then quantized. The quantization is a lossy coding, since process-related quantization errors occur. The quantization errors mean that the signals can not be accurately reconstructed after transmission and an unnatural stereo image is formed.

The center / side encoding has the effect, in addition to the data reducing effect, that when the left and right channels are very similar, the quantization error in both the left channel and the right channel coincides with the quantization error of the other channel is correlated, so that the quantization error takes place in the middle and there is covered by the useful signal a little or much better than in the uncorrelated case. However, as soon as the left and right channels are relatively dissimilar, due to the stereo effect, the useful signal will be either left or right, while the quantization error is correlated and more in the middle.

In order to obtain a further data volume reduction in the coding, the quantized middle / side signals are subsequently entropy-coded in the sense of a loss-free coding, for example by means of a Huffinan coding. By adding more information, such as scaling factors, a bit stream is formed from the quantized and entropy-coded center-to-side signals by means of a bit stream multiplexer which can be transmitted.

Scalable coding methods are particularly advantageous for stereo signals [ J. Li; Embedded Audio Coding (EAC) With Implicit Auditory Masking; ACM Multimedia 2002 ]. Scalable coding methods are designed such that the output-side bit stream has at least a first and a second scaling layer. The first scaling layer may differ from the second scaling layer or from any number of further scaling layers in the audio coding method itself, in the audio bandwidth, in the audio quality with respect to mono / stereo or in a combination of the quality criteria mentioned.

Scalable audio encoders for multichannel stereo transmission are often designed so that the mono signal, i.e., the first scaling layer, is the same. the center signal is used while in the other scaling layers the side channel is embedded. A decoder that is simply designed will take only the first scaling layer from the scaled bitstream and provide a mono signal. A stereo decoder also uses the side layer in addition to the center layer to provide a full bandwidth stereo signal.

A scalable stereo encoder that uses the center signal as the first scaling layer and the side signal in the other scaling layers has its best overall efficiency when there is a high similarity of the left channel to the right channel. For stereo channels that do not correlate or sudden changes in the characteristics of the two channels, the efficiency of a mid / side encoding is reduced.

The process of decoding a mid / side transmission is such that the received bit stream is divided by a demultiplexer into encoded quantized center / side signals and additional information. The entropy-coded quantized center-to-side signals are first entropy-decoded to obtain the quantized center-to-side signals, which are then inversely quantized. The decoded center / side signals have quantization errors introduced in the encoding and result in the signals for the left and right channels converted into the temporal representation after dematrixing and by means of a synthesis filter bank not being reconstructed in the original ratios can be

The object of the present invention is to achieve, for the application of the scalable coding according to the middle / side method, that spatial-related reproduction better obscures quantization errors and minimizes stereo imaging errors.

The object is achieved in that in the process of encoding the left channel and right channel are transformed and quantized by themselves and the middle / side processing only after the quantization takes place. The sum and subtraction is thus carried out with the already quantized signals of the left and right channels.

The invention is based on the finding that the effect of the quantization error in the center / side matrixing can be reduced if the matrixing is carried out after the quantization. This can be shown using the transfer equations.

The center signal is formed by the addition of the left and right channel, the side signal is formed by the difference. $$\begin{array}{l}\mathrm{M}\mathrm{=}\mathrm{0}\mathrm{.}\mathrm{5}\mathrm{R}\mathrm{+}\mathrm{0}\mathrm{.}\mathrm{5}\mathrm{L}\\ \mathrm{S}\mathrm{=}\mathrm{0}\mathrm{.}\mathrm{5}\mathrm{R}\mathrm{-}\mathrm{0}\mathrm{.}\mathrm{5}\mathrm{L}\end{array}$$

The recovery of the right and left channel is done with the operations: $$\begin{array}{l}\mathrm{R}\mathrm{=}\mathrm{M}\mathrm{+}\mathrm{S}\\ \mathrm{L}\mathrm{=}\mathrm{M}\mathrm{-}\mathrm{S}\end{array}$$

The quantization process is performed by the quantization function $$\mathrm{y}\mathrm{=}\mathrm{Q}\left(\mathrm{x}\right)\mathrm{described}\mathrm{,}$$

For conventional coding using quantization for the mid / side signals (M / S quantization), the transmission equations result: $$\begin{array}{c}\mathrm{R\; \text{'}}\mathrm{=}\mathrm{Q}\left(\mathrm{0}\mathrm{.}\mathrm{5}\mathrm{R}\mathrm{+}\mathrm{0}\mathrm{.}\mathrm{5}\mathrm{L}\right)\mathrm{+}\mathrm{Q}\left(\mathrm{0}\mathrm{.}\mathrm{5}\mathrm{R}\mathrm{-}\mathrm{0}\mathrm{.}\mathrm{5}\mathrm{L}\right)\\ \mathrm{L\text{'}}\mathrm{=}\mathrm{Q}\left(\mathrm{0}\mathrm{.}\mathrm{5}\mathrm{R}\mathrm{+}\mathrm{0}\mathrm{.}\mathrm{5}\mathrm{L}\right)\mathrm{-}\mathrm{Q}\left(\mathrm{0}\mathrm{.}\mathrm{5}\mathrm{R}\mathrm{-}\mathrm{0}\mathrm{.}\mathrm{5}\mathrm{L}\right)\end{array}$$

$$\mathrm{Lʹ}\mathrm{=}\mathrm{Q}\left(\mathrm{0}\mathrm{,}\mathrm{5}\mathrm{R}\mathrm{+}\mathrm{0}\mathrm{,}\mathrm{5}\mathrm{L}\right)$$

If only the mono signal is used for decoding then: $$\mathrm{R\; \text{'}}\mathrm{=}\mathrm{Q}\left(\mathrm{0}\mathrm{.}\mathrm{5}\mathrm{R}\mathrm{+}\mathrm{0}\mathrm{.}\mathrm{5}\mathrm{L}\right)$$

$$\mathrm{L\text{'}}\mathrm{=}\mathrm{Q}\left(\mathrm{0}\mathrm{.}\mathrm{5}\mathrm{R}\mathrm{+}\mathrm{0}\mathrm{.}\mathrm{5}\mathrm{L}\right)$$

$$\mathrm{S}\mathrm{=}\mathrm{0}\mathrm{,}\mathrm{5}\mathrm{Q}\left(\mathrm{R}\right)-\mathrm{0}\mathrm{,}\mathrm{5}\mathrm{Q}\left(\mathrm{L}\right)$$

The inventive optimization of the center / side stereophony using the quantization for the signals of the right and the left channel (R / L quantization) is as follows. The sum and difference signals are formed from the quantized R / L signals: $$\mathrm{M}\mathrm{=}\mathrm{0}\mathrm{.}\mathrm{5}\mathrm{Q}\left(\mathrm{R}\right)\mathrm{+}\mathrm{0}\mathrm{.}\mathrm{5}\mathrm{Q}\left(\mathrm{L}\right)$$

$$\mathrm{S}\mathrm{=}\mathrm{0}\mathrm{.}\mathrm{5}\mathrm{Q}\left(\mathrm{R}\right)-\mathrm{0}\mathrm{.}\mathrm{5}\mathrm{Q}\left(\mathrm{L}\right)$$

$$\begin{array}{c}\mathrm{Lʹ}\mathrm{=}\mathrm{0}\mathrm{,}\mathrm{5}\mathrm{Q}\left(\mathrm{R}\right)\mathrm{+}\mathrm{0}\mathrm{,}\mathrm{5}\mathrm{Q}\left(\mathrm{L}\right)-\end{array}\left(\mathrm{0}\mathrm{,}\mathrm{5}\mathrm{Q}\left(\mathrm{R}\right)\mathrm{-}\mathrm{0}\mathrm{,}\mathrm{5}\mathrm{Q}\left(\mathrm{L}\right)\right)$$

The insertion in equation (2) then yields: $$\begin{array}{c}\mathrm{R\; \text{'}}\mathrm{=}\mathrm{0}\mathrm{.}\mathrm{5}\mathrm{Q}\left(\mathrm{R}\right)\mathrm{+}\mathrm{0}\mathrm{.}\mathrm{5}\mathrm{Q}\left(\mathrm{L}\right)\mathrm{+}\mathrm{0}\mathrm{.}\mathrm{5}\mathrm{Q}\left(\mathrm{R}\right)\end{array}\mathrm{-}\mathrm{0}\mathrm{.}\mathrm{5}\mathrm{Q}\left(\mathrm{L}\right)$$

$$\begin{array}{c}\mathrm{L\text{'}}\mathrm{=}\mathrm{0}\mathrm{.}\mathrm{5}\mathrm{Q}\left(\mathrm{R}\right)\mathrm{+}\mathrm{0}\mathrm{.}\mathrm{5}\mathrm{Q}\left(\mathrm{L}\right)-\end{array}\left(\mathrm{0}\mathrm{.}\mathrm{5}\mathrm{Q}\left(\mathrm{R}\right)\mathrm{-}\mathrm{0}\mathrm{.}\mathrm{5}\mathrm{Q}\left(\mathrm{L}\right)\right)$$

This results for the optimization: $$\begin{array}{l}\mathrm{R\; \text{'}}\mathrm{=}\mathrm{Q}\left(\mathrm{R}\right)\\ \mathrm{L\text{'}}\mathrm{=}\mathrm{Q}\left(\mathrm{L}\right)\end{array}$$

$$\mathrm{Lʹ}\mathrm{=}\mathrm{0}\mathrm{,}\mathrm{5}\mathrm{Q}\left(\mathrm{R}\right)+\mathrm{0}\mathrm{,}\mathrm{5}\mathrm{Q}\left(\mathrm{L}\right)$$

If only the mono signal is used for decoding then: $$\mathrm{R\; \text{'}}\mathrm{=}\mathrm{0}\mathrm{.}\mathrm{5}\mathrm{Q}\left(\mathrm{R}\right)+\mathrm{0}\mathrm{.}\mathrm{5}\mathrm{Q}\left(\mathrm{L}\right)$$

$$\mathrm{L\text{'}}\mathrm{=}\mathrm{0}\mathrm{.}\mathrm{5}\mathrm{Q}\left(\mathrm{R}\right)+\mathrm{0}\mathrm{.}\mathrm{5}\mathrm{Q}\left(\mathrm{L}\right)$$

To evaluate the influence of the resulting quantization errors, a control of the system with stereo signals of the following form is considered: $$\begin{array}{l}\mathrm{xr}=a\mathrm{X}\\ \mathrm{X}1=\left(1-a\right)\mathrm{X}\end{array}$$

For a = 0, only the left channel is controlled, for a = 0.5 the left and right channels are equally controlled and for a = 1 only the right channel is controlled.

For conventional transmission using M / S quantization, the following output signals are obtained according to equation (4) for the input signals: $$\begin{array}{c}\mathrm{Xr\text{'}}\mathrm{=}\mathrm{Q}\left(\mathrm{0}\mathrm{.}\mathrm{5}\mathrm{X}\right)\mathrm{+}\mathrm{Q}\left(\mathit{a}\mathrm{X}\mathrm{-}\mathrm{0}\mathrm{.}\mathrm{5}\mathrm{X}\right)\\ \mathrm{Xl\text{'}}\mathrm{=}\mathrm{Q}\left(\mathrm{0}\mathrm{.}\mathrm{5}\mathrm{X}\right)\mathrm{-}\mathrm{Q}\left(\mathit{a}\mathrm{X}\mathrm{-}\mathrm{0}\mathrm{.}\mathrm{5}\mathrm{X}\right)\end{array}$$

For the optimization according to the invention using the R / L quantization, the following output signals are thus obtained: $$\begin{array}{l}\mathrm{Xr\text{'}}=\mathrm{Q}\left(a\mathrm{X}\right)\\ \mathrm{X}1\text{'}=\mathrm{Q}\left(\left(1-a\right)\mathrm{X}\right)\end{array}$$

At a value of a = 0.5, the results for the output signals are identical in both representations. The rule in practice, however, is that a assumes an arbitrary value between 0 and 1. Critical situations occur when a approaches 0 or 1. The one channel is then strongly controlled by the source signal, the other channel is weakly controlled.

To represent the quantization error, a quantizer with a size D quantization interval is assumed. The quantization error is denoted by d and can assume the values -D / 2 <d < D / 2 .

For the conventional application of M / S quantization, equation (7) yields: $$\begin{array}{c}\mathrm{Xr\text{'}}\mathrm{=}0.5\mathrm{X}\mathrm{+}\mathit{dm}+\left(\mathit{a}\mathrm{X}\mathrm{-}\mathrm{0}\mathrm{.}\mathrm{5}\mathrm{X}+\mathit{ds}\right)\\ \mathrm{Xl\text{'}}\mathrm{=}0.5\mathrm{X}\mathrm{+}\mathit{dm}-\left(\mathit{a}\mathrm{X}\mathrm{-}\mathrm{0}\mathrm{.}\mathrm{5}\mathrm{X}+\mathit{ds}\right)\end{array}$$

The quantization error of the center signal is dm, that of the side signal ds. There is a random relationship between dm and ds . The quantization error in the M / S quantization can take in the sum of values between - D and + D.

und für $$\begin{array}{l}a=0,5\\ \mathrm{Xrʹ}=0,5\mathrm{X}+\mathit{dm}+\mathit{ds}\\ \mathrm{Xlʹ}=0,5\mathrm{X}+\mathit{dm}-\mathit{ds}\end{array}$$

For the output signals resulting in a control with, for example $$\begin{array}{l}a=0\\ \mathrm{Xr\text{'}}=\mathit{dm}+\mathit{ds}\\ \mathrm{Xl\text{'}}=\mathrm{X}+\mathit{dm}-\mathit{ds}\end{array}$$

and for $$\begin{array}{l}a=0.5\\ \mathrm{Xr\text{'}}=0.5\mathrm{X}+\mathit{dm}+\mathit{ds}\\ \mathrm{Xl\text{'}}=0.5\mathrm{X}+\mathit{dm}-\mathit{ds}\end{array}$$

When a = 0 in the right channel, a quantization error is audible, although only the left channel having the signal. At a = 0.5, it can be seen that the quantization error occurs with in-phase and out-of-phase components. As a result, the quantization error becomes audible with a large stereo effect.

For the optimization according to the invention using the R / L quantization, the following relations result according to equation (8): $$\begin{array}{l}\mathrm{Xr\text{'}}=a\mathrm{X}+\mathit{dr}\\ \mathrm{X}1\text{'}=\left(1-a\right)\mathrm{X}+\mathit{dl}\end{array}$$

dr is the quantization error for the right channel, dl the quantization error for the left channel. For a quantization interval of size D , the quantization error d can assume the values - D / 2 < d < D / 2 , as already illustrated. In R / L quantization, the quantization errors do not add up. Thus the error remains in the range -D / 2 <d <D / 2.

$$\begin{array}{l}\mathrm{Xrʹ}=\mathit{dr}\\ \mathrm{X}1ʹ=\mathrm{X}+\mathit{dl}\end{array}$$

und für $$\begin{array}{l}a=0,5\\ \mathrm{Xrʹ}=0,5\mathrm{X}+\mathit{dr}\\ \mathrm{Xlʹ}=0,5\mathrm{X}+\mathit{dl}\end{array}$$

Im Vergleich zur herkömmlichen M/S-Quantisierung ist bei der R/L-Quantisierung nur ein Quantisierungsfehler möglich, der maximal halb so groß ist und keine gegenphasigen Komponenten aufweist, so dass das Nutzsignal den Quantisierungsfehler wesentlich besser verdeckt.For the output signals arise for $$a=0$$

$$\begin{array}{l}\mathrm{Xr\text{'}}=\mathit{dr}\\ \mathrm{X}1\text{'}=\mathrm{X}+\mathit{dl}\end{array}$$

and for $$\begin{array}{l}a=0.5\\ \mathrm{Xr\text{'}}=0.5\mathrm{X}+\mathit{dr}\\ \mathrm{Xl\text{'}}=0.5\mathrm{X}+\mathit{dl}\end{array}$$

Compared to conventional M / S quantization, only one quantization error is possible in R / L quantization, which is at most half as large and has no antiphase components, so that the useful signal obscures the quantization error much better.

embodiment

In Fig. 1, encoders and decoders are shown as an example of the application of the inventive principle of center / page formation after the quantization of the signals of the left and right channels. The description is limited to a two-channel transmission and encoding. However, the same principles can also be applied to multi-channel transmission and coding.

The left (10) and right channel (20) of an audio signal are first transformed from the time domain into the frequency domain. For this purpose, the known principle of the sliding modified cosine transform (200) is used for both audio channels. The spectral values of the left (11) and right (12) channels are quantized in the next step. The quantizer (300) is controlled by a quantization controller (500). The quantization can, as is known from other methods, be supported by a division into frequency bands. This division has the advantage that the quantization error is adapted to the spectral properties of the useful signal and thus is not audible so quickly for our hearing. The quantization is adapted to the modulation in the respective frequency band by a band for each band Scaling factor is determined. The quantization controller uses the left (10) and right (20) input channels to determine the scaling factors. A special feature of the quantization control in the new coding method is that the same scaling factor must be used for the left and right channel in order to enable the sum and difference formation in a linear number space. Apart from this constraint, various known methods can be used to determine the optimal scaling factors [ Marina Bosi and Karlheinz Brandenburg; Introduction to Digital Audio Coding and Standards; Springer Verlag 2002 ]. The quantization fulfills the function of a lossy reduction of the bits required for the coding.

The spectrally decomposed and quantized left (12) and right (22) channels are now fed to a center / side transformation stage (100) for converting left / right signals into center / side signals. Another data reduction took place in a further stage for lossless coding (400). This stage, which can be realized, for example, as usual in other coding methods with a Huffman coding, the center (40) and side signals (50) and the scaling factors (60) are supplied. The result is the coded signal (80).

The decoding of the encoded signal (80) is done by performing the steps in reverse order. The lossless decoding reconstructs the center (41) and side signals (51) as well as the scaling factors (61). In the next stage (101), the center and side signals are transformed back to left (13) and right (23) quantized signals. Thereafter, by means of the scaling factors (61), inverse quantization (301) is carried out to produce the original values of the spectral coefficients. The spectrally decomposed left (14) and right (15) signals are reset with the inverse modified discrete cosine transform (201) to the reconstructed signals for the left (15) and right (25) channels.

The present invention for minimizing the quantization errors also makes it possible in practice to make the generation of the bit stream more flexible. The encoded signal (80) can be scaled in size (bit rate). The bitstream contains the scaling factors, the center signal and the side signal. The bitrate can now be reduced in several ways. First, high-frequency components of the side signal can be omitted. Then, for example, the high frequency components of the center signal can be omitted. The unused scaling factors then do not need to be transmitted. In the next step, the low-frequency components of the side signal could then be reduced until, for example, the side signal no longer occurs in the bit stream. The quality of the stereo transmission can thus be transferred step by step into a mono transmission with decreasing spectral bandwidth.

Claims (3)

A method for scalable encoding of stereo signals, wherein the center / side encoding is applied and the left and right channel signals are transformed from the time domain to the frequency domain before the center / side processing and the signals are compressed to compress the data quantized after summation for summing and subtraction,
characterized,
that the signals of the left and right channels (10 and 20) after the transformation into the frequency domain (200) are quantized separately (300),
that the matrixing to form the center and side signals (100) with the already quantized signals of the left and right channels (12 and 22) takes place, and
that formed from the quantized signals of the left and right channel mid and side signals (40 and 50) are used in a further stage of the lossless encoding (400) for transmission in a coded signal (80). Method according to Claim 1, characterized in that the quantization of the signals of the left and right channels transformed into the frequency domain is supported by a division into frequency bands, that a scaling factor is determined for each frequency band that the scaling factors (60) are determined by the quantization controller (500 ) are derived from the signals of the left (10) and right (20) input channels, that the scaling factors for the left and right channels must be identical, and that the scaling factors are transmitted along with the center and side signals in the coded signal. A method according to claim 1 and 2, characterized in that the bit stream of the coded signal can be made flexible, so that a stepwise adaptation of the bit rate to the transmission conditions is advantageously possible.

Images