Note : Les descriptions sont présentées dans la langue officielle dans laquelle elles ont été soumises.
CA 02482427 2004-10-12
Apparatus and M,ethod for Coding a Time-na.scre.te Audio
signal and Apparatus and Method for Decoding
Coded Audio Data.
Description
The present invention relates to the audio coding/decoding,
and in particular to scalable coding/decoding algorithms
with a psychoacoustic first scaling layer and a second
scaling layer including ancillary audio data for lossless
decoding.
Modern audio coding methods, such as MPEG Layer3 (MP3) or
MPEG AAC, use transforms, such as the so-called modified
discrete cosine transfo= (MDCT), to obtain a block-wise
frequency representation of an audio signal. Such an audio
coder usually obtains a stream of time-discrete audio
samples. A stream of audio samples is windowed to obtain a
windowed block of for example 1,024 or 2,048 windowed audio
samples. For the windowing, various window functions are
employed, such as a sine window, etc.
The windowed time-discrete audio samples are then converted
to a spectral, representation by means of a filter bank. In
princip3.e, a Fourier transform, or a variety of the Fourier
transform for special reasons, such as a FFT or, as has
been set forth, a MDCT, may be employed for this. The block
of audio spectral values at the output of the filter bank
may then be processed further depending on demand. In the
above-referenced audio coders, a quantization of the audio
spectral values follows, wherein the quantization stages
are typically chosen so that the quantization noise
introduced by the quantizing lies below the psychoacoustic
masking threshold, i.e. is "masked away". The quantization
is a lossy coding. In order to obtain further data amount
reduction, the quantized spectral values are then entropy
coded for example by means of Huffman coding. By adding
side information, such as scale factors etc., a bit stream,
CA 02482427 2004-10-12
- 2 -
which may be stored or transmitted, is fozmed from the
entropy-coded quantized spectral values by means of a bit
stream multiplexer.
In the audio decoder, the bit stream is split up in coded
quantized spectral values and side information by means of
a bit stream de-multiplexer. The entropy-coded quantized
spectral values are at first entropy decoded to obtain the
quantized spectral values. The quantized spectral values
are then inversely quantized to obtain decoded spectral
values comprising quantization noise, which, however, lies
below the psychoacoustic masking threshold and will thus be
inaudible. These spectral values are then converted into a
temporal representation by means of a synthesis filter bank
to obtain time-discrete decoded audio samples. In the
synthesis filter bank, a transform algorithm inverse to the
transform algorithm has to be employed. Moreover, the
windowing has to be cancelled after the frequency-time
inverse or backward transform.
In order to achieve good frequency selectivity, modern
audio coders typically use block overlap. Such a case is
illustrated in Fig. 4a. At first for example 2,048 time-
discrete audio samples are taken and windowed by means of
means 402. The window embodying means 402 has a window
length of 2N samples and provides a block of 2N windowed
samples at the output side. In order to achieve window
overlap, by means of means 404, which is illustrated
separate from means 402 only for clarity reasons in Fig.
4a, a second block of 2N windowed samples is formed. The
2,046 samples fed to means 404, however, are not the time-
discrete audio samples immediately ensuing the first
window, but contain the second half of the samples windowed
by means 402 and additionally contain only 1,024 "new"
samples. The overlap is symbolically illustrated by means
406 in Fig. 4a, causing an overlapping degree of 50%. Both
the 2N windowed samples output by means 402 and the 2N
windowed samples output by means 404 are then subjected to
CA 02482427 2004-10-12
- ~ -
the MDCT algorithm by means of means 408 and 410,
respectively. Means 408 provides N spectral values for the
first window according to the known MDCT algorithm, whereas
means 410 also provides N spectral values, but for the
second window, wherein there is an overlap of 50% between
the first window and the second window.
In the decoder, the N spectral values of the first window,
as it is shown in Fig. 4b, are fed to means 412 performing
an inverse modified discrete cosine transform. The same
applies for the N spectral values of the second window.
These are fed to means 414 also performing an inverse
modified discrete cosine transform. Both means 412 and
means 414 each provide 2N samples for the first window and
2N samples for the second window, respectively.
In means 416, designated with TDAC (time domain aliasing
cancellation) in Fig. 4b, the fact is taken into account
that the two windows are overlapping. In particular, a
sample yl of the second half of the first window, i.e. with
an index N+k, is summed with a sample yy from the first
half of the second window, i.e. with an index k, so that N
decoded temporal samples result at the output side, i.e. in
the decoder.
It is to be noted that by the function of means 416, which
is also referred to as add function, the windowing
performed in the coder schematically illustrated by Fig. 4a
is taken into account somewhat automatically, so that in
the decoder illustrated by Fig. 4b no explicit "inverse
windowing" has to take place_
When the window function implemented by means 402 or 404 is
designated with w(k), wherein the index k represents the
time index, the condition has to be met that the squared
window weight w(k) added to the squared window weight
w(N4-k) together are 1, wherein k runs from 0 to N-1. When a
sine window is used, the window weights of which follow the
CA 02482427 2004-10-12
- 4 -
first half-wave of the sine function, this condition is
always met, since the square of the sine and the square of
the cosine for each angle together result in the value 1.
Disadvan'Cageous in the window method with ensuing MDCT
function described in Fig. 4a is the fact that the
windowing by multiplication of a tirne-discrete sample, when
it is thought of a sine window, it is achieved with a
floating-point number, since the sine of an angle between 0
and 180 degrees does not yield an integer, apart from the
angle 90 degrees. Even when integer time-discrete samples
are windowed, floating-point numbers result after the
windowing.
Therefore, even when no psychoacoustic coder is used, i.e,
when lossless coding is to be achieved, quantization is
necessary at the output of means 408 or 410 to be able to
perform reasonably manageable entropy coding.
When known transforms, aLs they have been described on the
basis of Fig. 4a, are to be employed for lossless audio
coding, either very fine quantization has to be employed to
be able to neglect the resulting error due to rounding the
floating-point numbers, or the error signal has to be
additionally coded for example in the tixne domain.
Concepts of the former kind, i.e. in which the quantization
is so finely adjusted that the resulting error due to the
rounding of the floating-point numbers is negligible, are
for example disclosed in the German patent DE 197 42 201
Cl. Here, an aud;.o signal is converted to its spectral
representation and quantized to obtain quantized spectral
values. The quantized spectral values zre then inversely
auantized, converted to the time domain, and compared with
the original audio signal. If the error, i.e. the error
between the original audio signal and the
quantized/inversely quantized audio signal, lies above an
error threshold, the quantizer is more finely adjusted in
CA 02482427 2004-10-12
- 5 -
feedback, and the comparison is performed again. The
iteration is terminated, when the error threshold is
underrun. The maybe still present residual signal is coded
with a time domain coder and written into a bit stream
including, apart from the time-domain-coded residual
signal, also coded spectral values having been quantized
according to the quantizer adjustments that were present at
the time of the cancellation of the iteration. It is to be
noted that the quantizer does not have to be controlled
from a psychoacoustic model, so that the coded spectral
values are typically quantized more accurately than this
would have to be due to the psychoacoustic model.
In the publication "A Design of Lossy and Lossless Scalable
Audio Coding", T. Moriya et al., Proc. ICASSP, 2000, a
scalable coder is described, which incl.udes e.g. an MPEG
coder as first lossy data compression module, which has a
block-wise digital signal :orm as input signal and
generates the compressed bit stream. In an also present
local decoder the coding is cancelled again, and a
coded/decoded signal is generated. This signal is compared
with the original input signal by subtracting the
coded/decoded signal from the original input signal. The
error signal is then fed to a second module, where a
lossless bit conversion is used_ This conversion has two
steps. The first step consists in a conversion from a two's
complement format to a presign-magnitude format. The second
step consists in a conversion from a vertical magnitude
sequence to a horizontal bit sequence in a processing
block. The lossless data conversion is executed to maximize
the number of zeros or to maximize the number of successive
zeros in a sequence, in order to achieve an as-good-as-
possible compression of the temporal error signal present
as a result of,digital numbers. This principle is based on
a bit slice arithmetic coding (BSAC) scheme illustrated in
the publication "Mu].ti-Layer Bit Sliced Bit Rate Scalable
Audio Coder", 103rd AES Convention, Preprint No. 4520,
1997.
CA 02482427 2009-11-02
- 6-
Disadvantageous in the above-described concepts is the fact
that the data for the lossless expansion layer, i.e. the
ancillary data required to achieve lossless decoding of the
audio signal has to be obtained in the time domain. This
means that complete decoding including a frequency/time
conversion is required to obtain the coded/decoded signal
in the time domain, so that by means of a sample-wise
difference formation between the original audio input
signal and the coded/decoded audio signal, which is lossy
due to the psychoacoustic coding, the error signal is
calculated. This concept is particularly disadvantageous in
that in the coder generating the audio data stream both
complete time/frequency conversion means, such as a filter
bank or e.g. a MDCT algorithm, is required for the forward
transform, and at the same time, only to generate the error
signal, a complete inverse filter bank or a complete
synthesis algorithm is required. The coder thus, in
addition to its inherent coder functionalities, also has to
contain the complete decoder functionality. If the coder is
implemented in software, both storage capacities and
processor capacities are required for this, leading to a
coder implementation with increased expenditure.
According to a first broad aspect of the invention, there
is provided an apparatus for coding a time-discrete audio
signal to obtain coded audio data, comprising: means for
providing a quantization block of spectral values of the
time-discrete audio signal quantized using a psychoacoustic
model; means for inversely quantizing the quantization
block to produce inversely quantized spectral values and
for rounding the inversely quantized spectral values to
obtain a rounding block of rounded inversely quantized
spectral values; means for generating an integer block of
integer spectral values using an integer transform
algorithm formed to generate the integer block of spectral
values from a block of integer time-discrete samples;
combination means for forming a difference block depending
on a spectral value-wise difference between the rounding
McCarthy Tetrault LLP DOCS #780712 v. 2
CA 02482427 2009-11-02
- 6a -
block and the integer block, to obtain a difference block
with difference spectral values; and means for processing
the quantization block and the difference block to generate
coded audio data including information on the quantization
block and information on the difference block.
According to a second broad aspect of the invention, there
is provided a method of coding a time-discrete audio signal
to obtain coded audio data, comprising: providing a
quantization block of spectral values of a time-discrete
audio signal quantized using a psychoacoustic model;
inversely quantizing the quantization block to produce
inversely quantized spectral values and rounding the
inversely quantized spectral values to obtain a rounding
block of rounded inversely quantized spectral values:
generating an integer block of integer spectral values
using an integer transform algorithm formed to generate the
integer block of spectral values from a block of integer
time-discrete samples; forming a difference block depending
on a spectral value-wise difference between the rounding
block and the integer block, to obtain a difference block
with difference spectral values; and processing the
quantization block and the difference block to generate
coded audio data including information on the quantization
block and information on the difference block.
According to a third broad aspect of the invention, there
is provided an apparatus for decoding coded audio data
having been generated from a time-discrete audio signal by
providing a quantization block of spectral values of the
time-discrete audio signal quantized using a psychoacoustic
model, by inversely quantizing the quantization block to
produce inversely quantized spectral values and rounding
the inversely quantized spectral values to obtain a
rounding block of rounded inversely quantized spectral
values, by generating an integer block of integer spectral
values using an integer transform algorithm formed to
generate the integer block of spectral values from a block
McCarthy Tetrault LLP DOCS #780712 v. 2
CA 02482427 2009-11-02
- 6b -
of integer time-discrete samples, and by forming a
difference block depending on a spectral value-wise
difference between the rounding block and the integer
block, to obtain a difference block with difference
spectral values, comprising: means for processing the coded
audio data to obtain a quantization block and a difference
block; means for inversely quantizing and rounding the
quantization block to obtain an integer inversely quantized
quantization block; means for spectral value-wise combining
the inversely quantized integer quantization block and the
difference block to obtain a combination block; and means
for generating a temporal representation of the time-
discrete audio signal using the combination block and using.
an integer transform algorithm that is inverse to the
integer transform algorithm that was used to generate the
integer block as aforesaid.
According to a fourth broad aspect of the invention, there
is provided a method of decoding coded audio data,
comprising: processing the coded audio data to obtain a
quantization block and a difference block; inversely
quantizing the quantization block and rounding same to
obtain an integer inversely quantized quantization block;
spectral value-wise combining the integer inversely
quantized quantization block and the difference block to
obtain a combination block; and generating a temporal
representation of the time-discrete audio signal using the
combination block and using an integer transform algorithm
that is inverse to another integer transform algorithm that
was used when generating the coded audio data.
The present invention intends to provide a less expensive
concept, by which an audio data stream may be generated,
which may be decoded in an at least almost lossless mar-ner.
The present invention is based on the finding that the
ancillary audio data enabling lossless decoding of the
McCarthy Tetrault LLP DOCS #780712 v. 2
CA 02482427 2009-11-02
- 7 -
audio signal may be obtained by providing a block of
quantized spectral values as usual and then inversely
quantizing it in order to have inversely quantized spectral
values, which are lossy due to the quantization by means of
a psychoacoustic model. These inversely quantized spectral
values are then rounded to obtain a rounding block of
rounded inversely quantized spectral values. As reference
for the difference formation, according to the invention,
an integer transform algorithm is used, which generates an
integer block of spectral values only comprising integer
spectral values from a block of integer time-discrete
samples. According to the invention, now the combination of
the spectral values in the rounding block and in the
integer block is performed spectral value-wise, i.e. in the
frequency domain, so that in the coder itself no synthesis
algorithm, i.e. an inverse filter bank or an inverse MDCT
algorithm, etc., is required. The combination block
comprising the difference spectral values only includes
integer values, which may be entropy coded in some known
manner, due to the integer transformation algorithm and the
rounded quantization values. It is to be noted that
arbitrary entropy coders may be employed for the entropy
coding of the combination block, such as Huffman coders or
arithmetic coders, etc.
For the coding of the quantized spectral values of the
quantization block, also arbitrary coders may be employed,
such as the known tools usual for modern audio coders.
It is to be noted that the inventive coding/decoding
concept is compatible with modern coding tools, such as
window switching, TNS, or center/side coding for multi-
channel audio signals.
In an embodiment of the present invention, a MDCT is
employed for providing a quantization block of spectral
values quantized using a psychoacoustic model. In addition,
McCarthy Tetrault LLP DOCS #780712 v. 2
CA 02482427 2009-11-02
- 8 -
in another embodiment a so-called IntMDCT as integer
transform algorithm is employed.
In an alternative embodiment of the present invention, it
can be done without the usual MDCT, and the IntMDCT may be
used as approximation for the MDCT, namely in that the
integer spectrum obtained by the integer transform
algorithm is fed to a psychoacoustic quantizer to obtain
quantized IntMDCT spectral values, which are then again
inversely quantized and rounded to be compared with the
original integer spectral values. In this case only a
single transform is required, namely the IntMDCT generating
integer spectral values from integer time-discrete samples.
Typically, processors work with integers, or each floating-
point number may be represented as an integer. If an
integer arithmetic is used in a processor, it can be done
without the rounding of the inversely quantized spectral
values, since due to the arithmetic of the processor
rounded values, namely within the accuracy of the LSB, i.e.
the least significant bit, are present anyway. In this
case, completely lossless processing is achieved, i.e.
processing within the accuracy of the used processor
system. Alternatively, however, rounding to a rougher
accuracy may be performed, in that the difference signal in
the combination block is rounded to an accuracy fixed by a
rounding function. Introducing rounding beyond the inherent
rounding of the processor system enables flexibility in so
far as to affect the "degree" of the losslessness of the
coding, in order to generate an almost lossless coder in
the sense of data compression.
The inventive decoder involves both the psychoacoustically
coded audio data and the ancillary audio data being
extracted from the audio data, being subjected to possibly
present entropy decoding, and then being processed as
follows. At first the quantization block in the decoder is
inversely quantized and rounded using the same rounding
McCarthy Tetrault LLP DOCS #780712 v. 2
CA 02482427 2009-11-02
- 9 -
function also employed in the coder, in order to be then
added to the entropy-decoded ancillary audio data. In the
decoder, then both a psychoacoustically compressed spectral
representation of the audio signal and a lossless
representation of the audio signal are present, wherein the
psychoacoustically compressed spectral representation of
the audio signal is to be converted to the time domain to
obtain a lossy coded/decoded audio signal, whereas the
lossless representation is converted in the time domain
using an integer transform algorithm inverse to the integer
transform algorithm to obtain a losslessly or, as it has
been set forth, almost losslessly coded/decoded audio
signal.
These and other objects and features of the present
invention will become clear from the following description
taken in conjunction with the accompanying drawings, in
which:
FIG. 1 is a block circuit diagram of preferred means for
processing time-discrete audio samples to obtain
integer values from which integer spectral values
can be ascertained;
FIG. 2 is a schematic illustration of the split-up of a
MDCT and an inverse MDCT in Givens rotations and
two DCT-IV operations;
FIG. 3 is a representation for the illustration of the
split-up of the MDCT with 50% overlap in
rotations and DCT-IV operations;
FIG. 4a is a schematic block circuit diagram of a known
coder with MDCT and 50 percent overlap;
FIG. 4b is a block circuit diagram of a known decoder for
decoding the values generated by FIG. 4a;
FIG. 5 is a principle block circuit diagram of a
preferred inventive coder;
McCarthy Tetrault LLP DOCS #780712 v. 2
CA 02482427 2008-10-08
CA 02482427 2004-10-12
- 10 -
Fig. 6 is a principle block circuit diagram of an
alternative inventively preferred coder;
Fig. 7 is a principle block circuit diagram of an
inventively preferred decoder;
Fig. 8a is a schematic illustration of a bit stream with
a first scaling layer and a second scaling layer;
Fig. 8b is a schematic illustration of a bit stream with
a first scaling layer and several further scaling
layers; and
Fig. 9 is a schematic illustration of binarily coded
difference spectral values for the illustration
of possible scalings with regard to the accuracy
(bits) of the difference spectral values and/or
with regard to the frequency (sample rate) of the
difference spectral values.
In the following, on the basis of Figs. 5 to 7, it is gone
into inventive coder circuits (Fig. 5 and Fig. 6) or an
inventively preferred decoder circuit (Fig. 7). The
inventive coder shown in Fig. 5 includes an input 50, to
which a time-discrete audio signal may be fed, as well as
an output 59, from which coded audio data may be output.
The time-discrete audio signal fed at the input 50 is fed
to means 52 for providing a quantization block, which
provides a quantization block of the time-discrete audio
signal at the output side, which comprises quantized
spectral values of the time-discrete audio signal 50 using
a psychoacoustic model 54. The inventive coder further
includes means for generating an integer block using an
integer transform algorithm 56, wherein the integer
algorithm is operative to generate integer spectral values
from integer time-discrete samples.
CA 02482427 2008-10-08
CA 02482427 2004-10-12
The inventive coder further includes means 58 for inversely
quantizing the quantization block output from means 52 and,
when another accuracy than the processor accuracy is
required, a rounding function. If it has to be gone up to
the accuracy of the processor system, as it has been set
forth, the rounding function already is inherently
contained in the inversely quantizing of the quantization
block, since a processor having an integer arithmetic is
incapable of providing non-integer values anyway. Means 58
thus provides a so-called rounding block including
inversely quantized spectral values, which are integer,
i.e. have been inherently or explicitly rounded. Both the
rounding block and the integer block are fed to combining
means providing a difference block with difference spectral
values, using difference formation, wherein the term
"difference block" is to imply that the difference spectral
values are values including differences between the integer
block and the rounding block.
Both the quantization block output from means 52 and the
difference block output from the difference formation means
5,7 are fed to processing means 60 performing for example
usual processing of the quantization block and also causing
for example entropy coding of the difference block. Means
60 for processing outputs coded audio data at the output
59, which contains both information on the quantization
block and includes information on the difference block.
In a first preferred embodiment, as shown in Fig. 6, the
time-discrete audio signal is converted to its spectral
representation by means of a MDCT and then quantized. The
means 52 for providing the quantization block thus consists
of the MOCT means 52a and a quantizer 52b.
In addition, it is preferred to generate the integer block
with an IntMDCT 56 as integer transform algorithm.
CA 02482427 2004-10-12
- 12 -
In Fig. 6, the processing means 60 shown in Fig. 5 is also
illustrated as bit stream coding means 60a for bit stream
coding the quantization block output by means 52b, as well
as by an entropy coder 60b for entropy coding the
difference block. The bit stream coder 60a outputs the
psychoacoustically coded audio data, whereas the entropy
coder 60b outputs an entropy-coded difference block. The
two output data of blocks 60a and 60b may be combined in a
bit stream in a suitable manner, which has the
psychoacoustically coded audio data as first scaling layer
and which has the additional audio data for lossiess
decoding as second scaling layer. The scaled bit stream
then corresponds to the coded audio data shown in Fig. 5 at
the output 52 of the coder.
In an alternative preferred embodiment, it may be done
without the MDCT block 52a of Fig. 6, as it is implied in
Fig. 5 by a dashed arrow 62. In this case the integer
spectrum provided by the integer transform means 56 is both
fed to the difference formation means 58 and to the
quantizer 52b of Fig. G. The spectral values generated by
the integer transform. are here in a way used as
approximation for a usual MDCT spectrum. This embodiment
has the advantage that only the IntMDCT algorithm is
present in the coder, and that not both the IntMDCT
algorithm and the MDCT algorithm have to be present in the
coder.
Again referring to Fig. 6, it is to be noted that the solid
blocks and lines illustrate a usual audio coder according
to one of the MPEG standards, whereas the dashed blocks and
lines illustrate the extension of such a usual MPEG cocder.
it is thus to be seen that no fundamental change of the
usual MPEG coder is necessary, but that the inventive
capture of the ancillary audio data for lossless coding by
means of an integer transform may be added without change
to the coder/decoder basic structure.
CA 02482427 2004-10-12
- 13 -
Fig. 7 shows a principle block circuit diagram of an
inventive decoder for decoding the coded audio data output
at the output 52 of Fig. 5. This is at first split up into
psychoacoustically coded audio data on the one hand and the
ancillary audio data on the other hand. The
psychoacoustically coded audio data is fed to a usual bit
stream decoder 70, whereas the ancillary audio data, when
having been entropy coded in the coder, is entropy coded by
means of an entropy coder 72. At the output of the bit
stream decoder 70 of Fig. 7, quantized spectral values are
present, which are fed to an inverse quantizer 74, which
may in principle be constructed identically with the
inverse quantizez in the means of Fig. 6. If an accuracy is
aimed at, which does not correspond to the processor
accuracy, in the decoder also rounding means 76 is
provided, which performs the same algorithm or the same
rounding function for mapping a real number to an integer,
as it may be also implemented in the means 58 of Fig. 6. In
a decoder-side combiner 78, the rounded inversely quantized
spectral values are preferably additively combined spectxal
vaZue-wise with the entropy-coded ancillary audio data, so
that in the decoder on the one hand inversely quantized
spectral values are present at the output of means 74 and
on the other hand integer spectral, values are present at
the output of the combiner 78_
The output-side spectral values of means 74 may then be
converted to the time domain by means of means 80 for
performing an inverse modified discrete cosine transform,
to obtain a lossy psychoacoustically coded and again
decoded audio signal. By means of means 82 for performing
an inverse integer MDCT (IntMDCT), the output signal of the
combiner 78 is also converted to its temporal
representation, in order to generate a losslessly
coded/decoded audio signal or, when a corresponding rougher
rounding has been employed, an almost losslessly coded and
again decoded audio signal.
CA 02482427 2004-10-12
- 14 -
In the following, it is gone into a special preferred
embodiment of the entropy coder 60b of Fig. 6. Since, in a
usual modern MPEG coder, several code tables selected
depending on average statistics of the quantized spectral
values are present, it is preferred to use the same code
tables or code books also for the entropy coding of the
difference block at the output of the combiner 58. Since
the magnitude of the difference block, i.e. of the residual
IntMDCT spectrum, depends on the accuracy of the
:.0 quantization, a codebook selection of the entropy coder 60b
may be performed without ancillary side information.
In a MPEG-2 AAC coder, the spectral coefficients, i.e. the
quantized spectral values, are grouped into scale factor
bands in the quantization block, wherein the spectral
values are weighted with a gain factor derived from a
corresponding scale factor associated with a scale factor
band. Since in this known coder concept a non-uniform
quantizer is used to quantize the weighted spectral values,
the size of the residual values, i.e. the spectral values
at the output of the combiner 58, does not only depend on
the scale factors but also on the quantized values
themselves. But since both the scale factors and the
quantized spectral values are contained in the bit stream,
which is generated by the means 60a of Fig_ 6, i_e. in the
psychoacoustically coded audio data, it is preferred to
perform a codebook selection in the coder depending on the
size of the difference spectral values and also to
ascertain, in the decoder, the code table used in the coder
on the basis of both the scale factors transmitted in the
bit stream and the quantized values. Since no side
information has to be transmitted for entropy coding the
difference spectral values at the output of the combiner
58, the entropy coding only leads to data rate compression,
without having to expend any signalization bits in the data
stream as side information for the entropy coder 60b.
CA 02482427 2004-10-12
- 15 -
In an audio coder according to the standard MPEG-2 AAC,
window switching is used to avoid pre-echoes in transient
audio signal areas. This technique is based on the
possibility to select window shapes individually in each
half of the MDCT window, and enables to vary the block size
in successive blocks. Similarly, the integer transform
algorithm in form of the IntMDCT, which is explained with
reference to Figs. 1 to 3, is executed to also use
different window shapes in windowing and in the time domain
aliasing section of the MDCT split-up. It is thus preferred
to use the same window decisions both for the integer
transform algorithm and for the transform algorithm for
generating the quantization block.
In a coder according to MPEG-2 AAC, also several further
coding tools exist, of which only TNS (temporal noise
shaping) and center/side (CS) stereo coding are to be
mentioned_ In TNS coding, just like in CS coding,
modification of the spectral values prior to the
quantization is performed. Consequently, the difference
between the IntMDCT values, i.e. the integer block, and the
quantized MDCT values increases. According to the
invention, the integer transform algorithm is formed to
admit both TNS coding and center/side coding also of
integer spectral values. The TNS technique is based on
adaptive forward prediction of the MDCT values over the
frequency. The same prediction filter calculated by a usual
TNS module in a signal-adaptive manner is preferably also
used to predict the integer spectral values, wherein, if
non-integer values arise thereby, downstream rounding may
be employed, in order to again generate integer values.
This rounding preferably takes place after each prediction
step. In the decoder, the original spectrum may again be
reconstructed by employing the inverse filter and the same
rounding function. Similarly, the CS coding may also be
applied to IntMDCT spectral values by applying rounded
Givens rotations with an angle of 7z/4, based on the lifting
CA 02482427 2004-10-12
- 16 -
scheme. Thereby, the original IntMDCT values in the decoder
may be reconstructed again.
2t is to be noted that the inventive concept in its
preferred embodiment with the IntMDCT as integer transform
algorithm may be applied to all MDCT-based hearing-adapted
audio coders. Only as an example, such coders are coders
according to MPEG-4 AAC Scalable, MPEG-4 AAC Low Oel.ay,
MPEG-4 BSAC, MPEG-4 Twin VQ, Dolby AC-3 etc.
In particular, it is to be noted that the inventive concept
is reversely compatible. The hearing-adapted coder or
decoder is not changed, but only extended. Ancillary
information for the lossless components may be transmitted
in the bit stream coded in a hearing-adapted manner in a
reversely compatible manner, such as in MPEG-2 AAC in the
field "Ancillary Data". The addition to the previous
hearing-adapted decoder drawn in a dashed manner in Fig. 7
may evaluate this ancillaxy data, and reconstruct, together
with the quantized MDCT spectrum, the IntMDCT spectrum in a
lossless manner from the hearing-adapted decoder.
The inventive concept of the psychoacoustic coding,
supplemented by lossless or almost lossless coding, is
particularly suited for the generation, transmission, and
decoding of scalable data streams. It is known that
scalable data streams include various scaling layers, at
least the lowest scaling layer of which may be transmitted
and decoded independently of the higher scaling layers.
Further scaling layers or enhancement layers are added to
the first scaling layer or base layer in a scalable
processing of data. A fully equipped coder may generate a
scalable data stream having a first scaling layer and in
principle having an arbitrary number of further scaling
layers. An advantage of the scaling concept is that, in the
case in which a broadband transmission channel is
available, the scaled data stream generated by the coder
may be transmitted completely, i.e. inclusive of all
CA 02482427 2004-10-12
- 17 -
scaling layers, via the broadband transmission channel. If,
however, only a narrowband transmission channel is present,
the coded signal may yet be transmitted via the
transmission channel, but only in form of the first scaling
layer or a certain number of further scaling la_vers,
wherein the certain number is smaller than the overall
number of scaling layers generated by the coder_ Of course,
the coder, adapted to a channel to which it is connected,
may already generate the base scaling layer or first
scaling layer and a number of further scaling layers
dependent on the channel.
on the decoder side, the scalable concept also has the
advantage that it is reversely compatible. This means that
a decoder that is only able to process the first scaling
layer simply ignores the second and further scaling layers
in the data stream and can generate a useful output signal.
zf, however,~the decoder is a typically more modern decodex
that is able to process several scaling layers from the
scaled data stream, this coder may be addressed with the
same data stream as a base decoder.
In the present invention, the basic scalability is that the
quantization block, i.e. the output of the bit stream coder
60a, is written to a first scaling layer 81 of Fig. 8,
which, when Fig. 6 is considered, includes
psychoacoustically coded data e.g. for a frame. The
preferably entropy-coded difference spectral values
generated by the combining means 58 are written into the
second scaling layer at simple scalability, which is
designated with 82 in Fig. 8a and thus includes the
ancilliary audio data for a frame_
If the transmission channel from the coder to the decoder
is a broadband transmission channel, both scaling layers 81
and 82 may be transmitted to the decoder. If, however, the
transmission channel is a narrowband transmission channel,
in which only the first scaling layer "fits", the second
CA 02482427 2004-10-12
- 18 -
scaling layer may simply be removed from the data stream
before the transmission, so that a decoder is only
addressed with the first scaling layer.
On the decoder side a "base decoder" that is only able to
process the psychoacoustically coded data may simply omit
the second scaling layer 82, as far it has received it via
a broadband transmission channel. If, however, the decoder
is a fully equipped decoder including both a psychoacoustic
decoding algorithm and an integer decoding algorithm, this
fully equipped decoder may take both the first scaling
layer and the second scaling layer for decoding to generate
a losslessly coded and again decoded output signal.
In a preferred embodiment of the present invention, as it
is schematically illustrated in Fig_ 8a, the
psychoacoustically coded data for a frame will again be in
a first scaling layer. The second scaling layer of Fig. 8a,
however, is now scaled more finely, so that from this
second scaling layer in Fig. 8a several scaling layers
arise, such as a (smaller) second scaling layer, a third
scaling layer, a fourth scaling layer, etc.
The difference spectral values output from the adder 58 are
particularly well suited for further subscaling, as it is
illustrated on the basis of Fig. 9. Fig. 9 schematically
illustrates binarily coded spectral values. Each row 90 in
Fig. 9 represents a binarily coded difference spectral
value. In Fig. 9 the difference spectral values are sorted
according to the frequen,cy, as it is implied by an arrow
91. A difference spectral value 92 thus has a higher
frequency than the difference spectral value 90. The first
column of the tablet in Fig. 9 presents the most
significant bit of a difference spectral value. The second
da.git represents the bit with a significance MSB-1. The
third column represents a bit with the significance MSB-2_
The last but second column represents a bit with the
significance LSB+2. The last but one column represents a
CA 02482427 2004-10-12
- 19 -
bit with the significance LSB+l. Finally, the last column
represents a bit with the significance LSB, i.e. the least
significant bit of a difference spectral value.
In a preferred embodiment of the present invention, an
accuracy scaling is made in that the e.g. 16 most
significant bits of a difference spectral value are taken
as second scaling layer, in order to then, if desired, be
entropy coded by the entropy coder 60b. A decoder using the
second scaling layer obtains difference spectral values
with an accuracy of 16 bits at the output side, so that the
second scaling layer, together with the first scaling
layer, provides a losslessly decoded audio signal in CD
quality. It is known that audio samples in CD quality with
a width of 16 bits are present.
If on the other hand an audio signal in studio quality is
fed to the coder, i.e. an audio signal with samples, with
each sample including 24 bits, the coder may further
generate a third scaling layer including the last eight
bits of a difference spectral value and also being entropy
coded depending on demand (means 60 of Fig. 6).
A fully equipped decoder obtaining the data stream with the
first scaling layer, the second scaling layer (16 most
significant bits of the difference spectral values), and
the third scaling layer (8 less significant bits of a
difference spectral value) may provide a losslessly
coded/decoded audio signal in studio quality, i.e. with a
word width of a sample of 24 bits present at the output of
the decoder, using all three scaling layers.
It is to be noted that in the studio area higher word
lengths of the samples are customary than in the consumer
area. In the consumer area the word width is 16 bits in an
audio CD, whereas in the studio area 24 bits or 20 bits are
employed.
CA 02482427 2004-10-12
- 20 -
Based on the concept of the scaling in the IntMDCT area, as
it has been set forth, thus all three accuracies (16 bits,
20 bits or 24 bits) or arbitrary accuracies scaled by
minimally 1 bit may be scalably coded.
Here, the audio signal, represented with 24 bit accuracy is
represented in the integer spectral region with the aid of
the inverse IntMDCT and scalably combined with a hearing-
adapted MDCT-based audiocoder output signal.
The integer difference values present for the lossless
representation are now not completely coded in a scaling
layer, but at first with lower accuracy. Only in a further
scaling layer are the residual values transmitted that
necessary for the exact representation. Alternatively
however, a difference spectral value could be represented
entirely, i.e. with for example 24 bits, also in a further
scaling layer, so that for decoding this further scaling
layer the underlying scaling layer is not reauired. This
scenario, however, altogether leads to a higher bit stream
size, but when the bandwidth of the transmission channel is
unproblematic may contribute to a simplification in the
decoder, since in the decoder scaling layers do then no
longer have to be combined, but always one scaling layer
alone is sufficient for decoding.
If for example the lower eight LSB, as it is illustrated in
Fig. 9, are not transmitted at first, a scalability between
24 bits and 16 bits is achieved.
For the inverse transform of the values transmitted with
lower accuracy into the time domain, the transmitted values
are preferably scaled back to the original region, for
example 24 bits, by multiplying them for exasaple by 2B. An
inverse IntMDCT is then applied to the correspondingly
scaled-back values.
CA 02482427 2004-10-12
- 21 -
In the inventive accuracy scaling in the frequency domain,
it is further preferred to also utilize the redundancy in
the LSBs. if an audio signal for example has very little
energy in the upper frequency domain, this also shows in
very small values in the IntMDCT spectrum, which are for
example significantly smaller than values (-128, ..., 127)
possible with for example 8 bits. This shows in a
compressibility of the LSB values of the IntMDCT spectrum.
Furthermore, it is to be noted that in very small
difference spectral values typically a number of bits from
MSs to MSB-1 are equal to zero, and that then the first,
leading 1 in a binarily coded difference spectral value
does not occur before a bit with a significance MSB-n-1. In
such a case, when a difference spectral value :in the second
scaling layer includes only zeros, entropy coding is
particularly well suited for the further data compression.
According to a further embodiment of the present invention,
for the second scaling layer 82 of Fig. Ba, a sample rate
scalability is preferred. A sample rate scalability is
achieved by the difference spectral values up to a first
cut-off frequency being contained in the second scaling
layer, as it is illustrated in Fig. 9 on the right, whereas
in a further scaling layer the difference spectral values
with a frequency between the first cut-off frequency and
the maximum frequency are contained. Of course, further
scaling may be performed, so that several scaling layers
are made from the entire frequency domain.
Yn a preferred embodiment of the px'esen'c invention, the
second scaling layer in Fig. 9 includes difference spectral
values up to a frequency of 24 kHz, corresponding to a
sample rate of 48 kHz. The third scaling layer then
contains the difference spectral values from 24 kHz to 48
kHz, corresponding to a sample rate of 96 kHz.
It is gurther to be noted that in the second scaling layer
and the third scaling layer not necessarily all bits of a
CA 02482427 2004-10-12
- 22 -
difference spectral value have to be coded. In a further
form of the combined scalability, the second scaling layer
could include bits MSB to MSB-X of the difference spectral
values up to a certain cut-off frequency. A third scaling
layer could then include the bits MSB to MSB-X of the
difference spectral values from the first cut-off frequency
to the maximum frequency. A fourth scaling layer could then
include the residual bits for the difference spectral
values up to the cut-off frequency. The last scaling layer
could then include the residual bits of the difference
spectral values for the upper frequencies. This concept
will lead to a division of the tablet in Fig. 9 into four
quadrants, each quadrant representing a scaling layer.
In the scalability in frequency, in a preferred embodiment
of the present invention, a scalability between 48 kHz and
96 kHz sample rate is described. The 96 kHz sample signal
is at first only coded half in the IntMDCT area in the
lossless extension layer and transmitted. If the upper part
is not transmitted in addition, it is assumed zero in the
decoder. In the inverse IntMDCT (same length as in the
coder), then a 96 kHz signal arises, which does not contain
energy in the upper frequency domain and may thus be
subsampled on 48 kHz without quality losses.
The above scaling of the difference spectral values in
quadrants of Fig. 9 with fixed boundaxies is favorable
regarding the size of the scaling layers, because in a
scaling layer in fact only e.g. 16 bits or 8 bits or the
spectral values up to the cut-off frequency or above the
cut-off frequency have to be contained.
An alternative scaling is to somewhat "soften" the quadrant
boundaries in Fig. 9. In the example of the frequency
scalability this would mean not to apply a so-called
"brickwall low pass" in that the difference spectral values
before a cut-off frequency are unchanged and are zero after
the cut-off frequency. Instead, the difference spectral
CA 02482427 2004-10-12
- 23 -
values could also be filtered with an arbitrary low pass
already somewhat impeding the spectral values below the
cut-off frequency, but, above the cut-off frequency,
leading to here also still being energy, although the
difference spectral values are decreasing in energy. In a
so-generated scaling layer, then also spectral values above
the cut-off frequency are contained. Since these spectral
values, however, are relatively small, they are efficiently
codable by entropy coding. The highest scaling layer would
in this case have the difference between the complete
difference spectral values and the spectral values
contained in the second scaling layer.
The accuracy scaling may also somewhat be softened
similarly. The first scaling layer may also have spectral
values with e_g. more than 16 bits, wherein the next
scaling layer then still has the difference. Generally
speaking, the second scaling layer thus has the difference
spectral values with lower accuracy, whereas in the next
scaling layer the rest, i.e. the difference between the
complete spectral values and the spectral values contained
in the second scaling layer, is transmitted. With this,
variable accuracy reduction is achieved.
The inventive method for coding or decoding is preferably
stored on a digital storage medium, such as a floppy disc,
with electronically readable control signals, wherein the
control signals may cooperate with a programmable computer
system so that the coding and/or decoding method may be
executed. In other words, a computer program product with a
program code stored on a machine-readable carrier for
performing the coding method and/or the decoding method is
present, when the program product is executed on a
computer. The inventive method may be realized in a
computer program wi.th a program code for performing the
inventive methods, when the program is executed on a
computer.
CA 02482427 2004-10-12
- 24 -
In the following, as an example for an integer transform
algorithm, it is gone into the IntMDCT transform algorithm
described in "Audio Coding Based on Integer Transforms"
lllth AES convention, New York, 2001. The IntMDCT is
na:ta.cularly favorable, since it has the attractive
properties of the MDCT, such as good spectral
representation of the audio signal, critical sampling, and
block overlap. A good approximation of the MDCT by an
InrMDCT also enables to use only one transform algorithm in
the coder shown in Fig. 5, as it is illustrated by an arrow
62 in Fig. 5. On the basis of Figs. 1 to 4, the substantial
properties of this special form of an integer transform
algorithm are explained.
Fig. 1 shows an overview diagram for the inventively
preferred apparatus for processing time-discrete samples
representing an audio signal, in order to obtain integer
values based on which the Int-MDCT integer transform
algorithm works. The time-discrete samples are windowed and
optionally converted to a spectral representation by the
apparatus shown in Fig. 1. The time-discrete samples fed to
the apparatus at an input 10 are windowed with a window w
with a length corresponding to 2N time-discrete samples, to
achieve integer windowed samples at an output 12, which are
suited to be converted to a spectral representation by
means of a transform and in particular the means 14 for
executing an integer DCT. The integer bCT is formed to
generate N output values from N input values, which is in
contrast to the MDCT function 408 of Fig. 4a, which only
generates N spectral values from 2N windowed samples due to
the MDCT equation.
For windowing the time-discrete samples, at first two time-
discrete samples are selected in means 16, which together
represent a vector of time-discrete samples. A time-
discrete sample selected by means 16 lies in the first
quarter of the window. The other time-discrete sample lies
in the second quarter of the window, as it is explained in
CA 02482427 2004-10-12
- 25 -
still greater detail on the basis of Fig. 3. To the vector
generated by means 16 is now a rotation matrix of the
dimension 2 x 2 is applied, wherein this operation is not
performed immediately, but by means of several so-called
lifting matrices.
A lifting matrix has the property of only comprising one
element dependent on the window w and being unequal "1" or
,. 0 A. .
The factorization of wavelet transforms into lifting steps
is illustrated in the publication "Factoring Wavelet
Transforms Into Lifting Steps", Ingrid Daubechies and wim
Sweldens, preprint, Bell Laboratories, Lucent Technologies,
1996. In general, a lifting scheme is a simple relation
between perfectly reconstructed filter pairs having the
same low-pass or high-pass filter. Each pair of
complementary filters may be factorized into lifting steps.
This applies in particular to Givens rotations. Consider
the case in which the poly-phase matrix is a Givens
rotation. Then, the following applies:
cos a- sin a_ 1l~ 1
'õ~a' (1)
(sina casa )-~0 1) (sina 01) 1 0 "1
Each of the three lifting matrices to the right of the
equality sign has the value "1" as main diagonal elements.
Furthermore, in each lifting matrix an element not on the
main diagonal equals 0, and an element not on the main
diagonal is dependent on the rotation angle CL.
The vector is now multiplied by the third lifting matrix,
i.e. the lifting matrix on the far right in the above
equation, to obtain a first result veczor. This is
illustrated in Fig. 1 by means 18. Now the first result
vector is rounded with an arbitrary rounding function
mapping the set of real numbers to the set of integers, as
it is illustrated in Fig. 1 by means 20. At the output of
CA 02482427 2004-10-12
- 26 -
means 20, a rounded first result vector is obtained. The
rounded first result vector is now fed to means 22 for
multiplying it by the center, i.e. second, lifting matrix,
to obtain a second result vector, which is again, rounded in
means 24, to obtain a rounded second result vector. The
rounded second result vector is now fed to means 26 for
multiplying it by the lifting matrix set forth on the left
in the above equation, i.e. the first one, to obtain a
third result vector which is in the end still rounded by
means of means 28 to obtain integer windowed samples in the
end at the output 12, which now, when a spectral
representation thereof is desired, have to be processed by
means 14 to obtain integer spectral values at a spectral
output 30.
Preferably, means 14 is embodied as integer OCT.
The discrete cosine transform according to type 4 (DCT-IV)
with a length N is given by the following equation:
FW2 N-'~
XrX(k)co(2k +1)(2m +1)( 2 )
k_o 4N
The coefficients of the DCT-Iv form an orthonormal N x N
matrix. Each orthogonal N x N matrix may be split up into N
(N-1) /2 Givern3 rotations, as it is explained in the
publication P. P. Vaidyanathan, "Multirate Systems And
Filter Banks", Prentice Hall, Englewood Cliffs, 1993. It is
to be noted that there are also further split-ups.
With reference to the classificat.ions of the various DCT
algorithms, reference is to be made to H. S. Malvar,
"Signal Processing With Lapped Transforms", Artech House,
1992. In general, the DCT algorithms differ by the kind of
their basis functions. While the DCT-IV, which is preferred
here, includes non-symmetrical basis functions, i.e. a
cosine quarter wave, a cosine 3/4 wave, a cosine 5/4 wave,
a cosine 7/4 wave, etc., the discrete cosine transform e.g.
CA 02482427 2004-10-12
- 27 -
of the type II (DCT-II) has axis-symmetrical and point-
symmetrical basis functions. The 0th basis function has a
DC component, the first basis function is half a cosine
wave, the second basis function is a whole cosine wave,
etc. Due to the fact that the OCT-II particularly takes the
DC component into account, it is used in the video coding,
but not in the audio coding, since in the audio coding in
contrast to the video coding the DC component is
irrelevant.
in the following, it is gone into how the rotation angle a
of the Givens rotation depends on the window function.
A MDCT with a window length of 2N may be reduced to a
discrete cosine transform of type IV with a length N. This
is achieved by the TDAC operation being performed
explicitly in the time domain and the OCT-IV then being
applied. With a 50% overlap, the left half of the window
for a block t overlaps with the right half of the preceding
block, i.e. the block t-1. The overlapping part of two
sucCessive blocks t-1 and t is preprocessed in the time
domain, i.e. before the transform, i.e. between the input
10 and the output 12 of Fig. 1, as follows:
x~(k) ~'(i +k) -w(2 -1-k) +k)
/ l (3)
(3F_'(N-1-k))
,wlz -1-k) w(i +k) )~x~l~-'1-k
The values designated with the tilde are the values at the
output 12 of Fig_ 1, whereas x values designated without
tilde in the above equation are the values at the input 10
or behind the means 16 for selecting. The running index k
runs from 0 to N/2-1, while w represents the window
function.
From the TDAC condition for the window function w, the
following connection applies:
CA 02482427 2004-10-12
- 28 -
2 -rk}~+wl 2 --3-k~~ =1 (4)
For certain angles ak, k = 0, ..., N/2-1, this
preprocessing in the time domain may be written as Givens
rotation, as it has been explained.
The angle a of the Givens rotation depends on the window
function w as follows:
a - arctan [w (N/2-1-k) / w(N/2 + k) ] (5)
It is to be noted that arbitrary window functions w may be
employed as long as they meet this TDAC condition.
in the following, on the basis of Fig. 2, a cascaded coder
and decoder is described. The time-discrete samples x(0) to
x(2N-1) "windowed" together by a window are at first
selected by means 16 of Fig. 1 such that the sample x(0)
and the sample x(N-1), i.e_ a sample from the first quarter
of the window and a sample from the second quarter of the
window, are selected to form the vector at the output of
means 16. The crossing arrows schematically illustrate the
lifting multiplications and ensuing roundings of means 18,
20 or 22, 24 or 26, 28, in order to obtain the integer
windowed samples at the input of the DCT-IV blocks_
When the first vector is processed as described above, also
a second vector is selected from the samples x(N/2-1) and
x(N/2), i.e. again a sample from the first quarter of the
window and a sample from the second quarter of the window,
and again processed by the algorithm described in Fig. 1.
zn analogy therewith, all other sample pairs from the first
and second quarters of the window are treated. The same
processing is performed for the third and fourth quarters
of the first wzndow. At the output 12, now N windowed
integer samples are present, which are now fed to a DCT-IV
CA 02482427 2004-10-12
- 29 -
transform, as it is illustrated in Fig. 2. In particular,
the integer windowed samples of the second and third
quarters are fed to a DCT. The windowed integer samples of
the first quarter of the window are processed, together
with the windowed integer samples of the fourth quarter of
the preceding window, into a preceding DCT-IV. In analogy
therewith, the fourth quarter of the windowed integer
samples in Fig. 2, together with the first quarter of the
next window, is fed to a DCT-IV transform. The center
integer DCT-IV transform 32 shown in Fig. 2 now provides N
integer spectral values y(O) to yfN-1). These integer
spectral values may now for example simply be entropy
coded, without an intervening quantization being required,
since the windowing and transform provide integer output
values.
In the right half of Fig. 2, a decoder is illustrated. The
decoder including inverse transform and "inverse windowing"
works inversely to the coder. It is known that for the
inverse transform of a DCT-IV, an inverse DCT-IV may be
used, as it is illustrated in Fig. 2. The output values of
the decoder DCT-IV 34 are now, as it is illustrated in Fig.
2, inversely processed with the corresponding values of the
preceding transform or the following transform, in order to
generate again time-discrete audio samples x(0) to x(2N-1)
from the integer windowed samples at the output of means 34
or the preceding and following cransform.
The output-side operation takes place by an inverse Givens
rotation, i.e. such that the bZocks 26, 28 or 22, 24 or 18,
20 are passed in the opposite direction. This is to be
illustrated in greater detail on the basis of the second
lifting matrix of equation 1. When (in the coder) the
second result vector is formed by multiplication of the
rounded first result vector by the second lifting matrix
{means 22), the following term results:
(x, y) H(z, y+ xsina) (6)
CA 02482427 2004-10-12
- 30 -
The vaJ.ues x, y on the right side of equation 6 are
integers. This however does not apply for the value x sin
a. Here, the rounding function r has to be int.roduced, as
it is illustrated in the following equation.
(x,y) H (x,y+r(xstna)) (7)
This operation executes means 24.
The inverse mapping (in the decoder) is defined as follows:
(x', y') I-H (x',y'-r(x'si.na)) (8)
Due to the minus sign in front of the rounding operation,
it becomes apparent that the integer approximation of the
lifting step may be reversed, without introducing an error.
The application of this approximation to each of the three
lifting steps leads to an integer approximation of the
Givens rotation. The rounded rotation (in the coder) may be
reversed (in the decoder), without introducing an error,
namely by passing the inverse rounded lifting steps in
reversed order, i.e. when in decoding the algoxithm of Fig.
1 is performed from bottom to top.
If the rounding function r i.s point-symanetrical, the
inversed rounded rotation is identical to the rounded
rotation with the angle -a, and reads as follows:
cosa sina
(9)
-sina cosa
The lifting matrices for the decoder, i.e. for the inverse
Givens rotation, in this case immediately result from
equation (1), by simply replacing the term "sin a" by the
texm "-sin a".
In the following, on the basis of Fig. 3, the split-up of a
usual MDCT with overlapping windows 40 to 46 is set forth
CA 02482427 2004-10-12
- 31 -
once again. The windows 40 to 46 each overlap 50%. Per
window, at first Givens rotations within the first and
second quarters of a window or within the third and fourth
quarters of a window are executed, as it is schematically
illustrated by the arrows 48. Then, the rotated values,
i.e. the windowed integer samples, are fed to an N-to-N DCT
such that always the second and third quarters of a window
or the fourth and first quarters of a successive window are
together converted to a spectral representation by means of
a DCT-IV algorithm.
Therefore, the usual Givens rotations are split up into
lifting matrices, which are executed sequentially, wherein
after each lifting matrix multiplication a rounding step is
introduced such that the floating-point numbers are rounded
immediately after their development such that before each
multiplication of a result vector by a lifting matrix the
result vector has only integers.
The output values always stay integer, it being preferred
to also use integer input values. This does not represent a
limitation, since any exemplary PCM samples, as they are
stored on a CD, are integer number values the value range
of which varies depending on bit width, i.e. depending on
whether the time-discrete digital input values are 16-bit
values or 24-bit values_ Nevertheless, as it has been set
forth, the entire process is invertible by executing the
inverse rotations in reversed order. Thus, an integer
approximation of the MDCT with perfect reconstruction
exists, namely a lossless transform.
The transform shown provides integer output values instead
of floating-point values. Tt provides a perfect
reconstruction, so that no error is introduced when a
forward and then a backward transform are executed_ The
transform, according to a preferred embodiment of the
present invention, is a replacement for the modified
discrete cosine transform_ Other transform methods may,
CA 02482427 2004-10-12
- 32 -
however, also be executed in an integer manner, as long as
a split-up into rotations and a split-up of the rotations
into lifting steps is possible.
The integer MDCT has most of the favorable properties of
the MDCT. It has an overlapping structure, whereby better
frequency selectivity than in non-overlapping block
transforms is obtained. Due to the TDAC function, which is
already taken into account when windowing p--ior to the
transform, critical sampling is maintained so that the
overall number of spectral values representing an audio
signal equals the overall number of input samples.
Compared with a normal MDCT providing floating-point
samples, in the described preferred integer transform, it
shows that only in the spectral region in which there is
little signal level the noise is increased in comparison
with the normal MDCT, whereas this noise increase does not
make itself felt at significant signal levels. For this,
the integer processing lends itself for an efficient
hardware implementation, since only multiplication steps
are used, which may easily be split up into shift/add
steps, which may be implemented in hardware easily and
quickly_ Of course, a software implementation is also
possible.
The integer transform provides a good spectral
representation of the audio signal and yet remains in the
area of integers. When it is applied to tonal parts 'of an
audio signal, this results in good energy concentration.
With this, an efficient lossless coding scheme may be built
up by simply cascading the windowing/transform illustrated
in Fig. I with an entropy coder. In particular, stacked
coding using escape values, as it is employed in MPEG AAC,
is favorable. It is preferred to scale down all values by a
certain power of two until they fit in a desired code
table, and then code the omitted least significant bits in
addition. in comparison with the alternative of the use of
CA 02482427 2004-10-12
- 33 -
larger code tables, the alternative described is more
favorable with regard to the storage consumption for
storing the code tables. An almost lossless coder could
also be obtained by simply omitting certain of the least
significant bits.
In particular for tonal signals, entropy coding of the
integer spectral values enables high coding gain. For
transient parts of the signal, the coding gain is low,
namely due to the flat spectrum of transient signals, i.e.
due to a small number of spectral values equal to or almost
0. As it is described in J. Herre, J. D. Johnston:
"Enhancing the Performance of Perceptual Audio Coders by
Using Temporal Noise Shaping (TNS)" 1015L AES Convention,
Los Angeles, 1996, preprint 4384, this flatness may however
be used by using a linear prediction in the frequency
domain. An alternative is a prediction with open loop.
Another alternative is the predictor with closed loop. The
first alternative, i.e. the predictor with open loop, is
called TNS. The quantization after the prediction leads to
adaptation of the resulting quantization noise to the
temporal structure of the audio signal and thus prevents
pre-echoes in psychoacoustic audio coders. For lossless.
audio coding, the second alternative, i.e. with a predictor
with closed loop, is more suited, since the prediction with
closed loop allows accurate reconstruction of the input
signal. When this technique is applied to a generated
spectrum, a rounding step has to be performed after each
step of the prediction filter in order to stay in the area
of the integers. By using the inverse filter and the same
rounding funetion, the original spectrum may accurately be
produced.
In order to take advantage of the redundancy between two
channels for data reduction, also center-side coding may be
employed in a lossless manner, when a rounded rotation with
an angle a/4 is used. Xn comparison with the alternative
of calculating the sum and difference of the left and right
CA 02482427 2004-10-12
- 34 -
channel of a stereo signal, the rounded rotations have the
advantage of the energy maintenance. The use of so-called
joint stereo codzng techniques may be switched on or off
for each band, as it is also performed in the standard MPEG
A.AC_ Further rotation angles may also be taken into account
to be able to reduce redundancy between two channels more
flexibly.
