Indian Patents
Title of Invention

DEVICE AND METHOD FOR POSTPROCESSING SPECTRAL VALUES AND ENCODER AND DECODER FOR AUDIO SIGNALS
Abstract For postprocessing spectral values which are based on a first transformation algorithm for converting the audio signal into a spectral representation, first a sequence of blocks of the spectral values representing a sequence of blocks of samples of the audio signal are provided. Here- upon, a weighted addition of spectral values of the se- quence of blocks of spectral values is performed in order to obtain a sequence of blocks of postprocessed spectral values, wherein the combination is performed such that for calculating a postprocessed spectral value for a frequency band and a time duration a spectral value of the sequence of blocks for the frequency band and the time duration and a spectral value for another frequency band or another time duration are used, wherein the combination is further per- formed such that such weighting factors are used that the postprocessed spectral values are an approximation to the spectral values as they are obtained by converting the au- dio signal into a spectral representation using a second transformation algorithm which is different from the first transformation algorithm. The postprocessed spectral values are in particular used for a difference formation within a scalable encoder or for an addition within a scalable decoder, respectively.
Full Text DEVICE AND METHOD FOR POSTPROCESSING SPECTRAL VALUES AND
ENCODER AND DECODER FOR AUDIO SIGNALS
Description
1. Field of the invention:
The present invention relates to audio encoding/decoding
and in particular to scalable encoder/decoder concepts hav-
ing a base layer and an extension layer.
2. Description of the related art:
Audio encoders/decoders have been known for a long time. In
particular audio encoders/decoders operating according to
the standard ISO/IEC 11172-3, wherein this standard is also
known as the MP3 standard, are referred to as transforma-
tion encoders. Such an MP3 encoder receives a sequence of
time samples as an input signal which are subjected to a
windowing. The windowing leads to sequential blocks of time
samples which are then converted into a spectral represen-
tation block by block. According to the MP3 standard, here
a conversion is performed with a so-called hybrid filter
bank. The first stage of the hybrid filter bank is a filter
bank having 32 channels in order to generate 32 subband
signals. The subband filters of this first stage comprise
overlapping passbands, which is why this filtering is prone
to aliasing. The second stage is an MDCT stage to divide
the 32 subband signals into 576 spectral values. The spec-
tral values are then quantized considering the psychoacous-
tic model and subsequently Huffman encoded in order to fi-
nally obtain a sequence of bits including a stream of Huff-
man code words and side information necessary for decoding.
On the decoder side, the Huffman code words are then calcu-
lated back into quantization indices. A requantization
leads to spectral values which are then fed into a hybrid
synthesis filter bank which is implemented analog to the
analysis filter bank to again obtain blocks of time samples
of the encoded and again decoded audio signal. All steps on

the encoder side and on the decoder side are presented in
the MP3 standard. With regard to the terminology it is
noted that in the following reference is also made to an
"inverse quantization". Although a quantization is not in-
vertible, as it involves an irretrievable data loss, the
expression inverse quantization is often used, which is to
indicate a requantization presented before.
Also an audio encoder/decoder algorithm called AAC (AAC =
Advanced Audio Coding) is known in the art. Such an encoder
standardized in the international standard ISO/IEC 13818-7
again operates on the basis of time samples of an audio
signal. The time samples of the audio signal are again sub-
jected to a windowing in order to obtain sequential blocks
of windowed time samples. In contrast to the MP3 encoder in
which a hybrid filter bank is used, in the AAC encoder one
single MDCT transformation is performed in order to obtain
a sequence of blocks of MDCT spectral values. These MDCT
spectral values are then again quantized on the basis of a
psychoacoustic model and the quantized spectral values are
finally Huffman encoded. On the decoder side processing is
correspondingly. The Huffman code words are decoded and the
quantization indices or quantized spectral values, respec-
tively, obtained therefrom are then requantized or in-
versely quantized, respectively, to finally obtain spectral
values that may be supplied to an MDCT synthesis filter
bank in order to finally obtain encoded/decoded time sam-
ples again.
Both methods operate with overlapping blocks and adaptive
window functions as described in the experts publication
"Codierung von Audiosignalen mit uberlappender Transforma-
tion und adaptiven Fensterfunktionen", Bernd Edler, Fre-
quenz, vol. 43, 1989, pp. 252-256.
In particular when transient areas are determined in the
audio signal, a switch is performed from long window func-
tions to short window functions in order to obtain a re-

duced frequency resolution in favor of a better time reso-
lution. A sequence of short windows is introduced by a
start window and a sequence of short windows is terminated
by stop a window. Thereby, a gapless transition between
overlapping long window functions to overlapping short win-
dow functions may be achieved. Depending on the implementa-
tion, the overlapping area with short windows is smaller
than the overlapping area with long windows, which is rea-
sonable with regard to the fact that transient signal por-
tions are present in the audio signal, does not necessarily
have to be the case, however. Thus, sequences of short win-
dows as well as sequences of long windows may be imple-
mented with an overlap of 50 percent. In particular with
short windows, however, for improving the encoding of tran-
sient signal portions, a reduced overlap width may be se-
lected, like for example only 10 percent or even less in-
stead of 50 percent.
Both, in the MP3 standard and also in the AAC standard the
windowing exists with long and short windows and the start
windows or stop windows, respectively, are scaled such that
in general always the same block raster may be maintained.
For the MP3 standard this means, that for each long block
576 spectral values are generated and that three short
blocks correspond to one long block. This means, that one
short block generates 192 spectral values. With an overlap
of 50 percent, for windowing thus a window length of 1152
time samples is used, as due to the overlap and add princi-
ple of a 50 percent overlap two blocks of time samples al-
ways lead to one block of spectral values.
Both with MP3 encoders and also with AAC encoders, a lossy
compression takes place. Losses are introduced by a quanti-
zation of the spectral values taking place. The spectral
values are in particular quantized so that the distortions
introduced by the quantization also referred to as quanti-
zation noise have an energy which is below the psychoacous-
tic masking threshold.

The coarser an audio signal is quantized, i.e. the greater
the quantizer step size, the higher the quantization noise.
On the other hand, however, for a coarser quantization a
smaller set of quantizer output values is to be considered,
so that values quantized coarser may be entropy encoded us-
ing less bits. This means, that a coarser quantization
leads to a higher data compression, however simultaneously
leads to higher signal losses.
These signal losses are unproblematic if they are below the
masking threshold. Even if the psychoacoustic masking
threshold is only exceeded slightly, this may possibly not
yet lead to audible interferences for unskilled listeners.
Anyway, however, an information loss takes place which may
be undesired for example due to artifacts which may be au-
dible in certain situations.
In particular with broadband data connections or when the
data rate is not the decisive parameter, respectively, or
when both broadband and also narrowband data networks are
available, it may be desirable to have not a lossy but a
lossless or almost lossless, compressed presentation of an
audio signal.
Such a scalable encoder schematically illustrated in Fig. 7
and an associated decoder schematically illustrated in
Fig. 8 are known from the experts publication "INTMDCT - A
Link Between Perceptual And Lossless Audio Coding", Ralf
Geiger, Jiirgen Herre, Jürgen Roller, Karlheinz Brandenburg,
Int. Conference on Acoustics Speech and Signal Processing
(ICASSP), 13 - 17 May, 2002, Orlando, Florida. A similar
technology is described in the European Patent EP 1 495 464
Bl. The elements 71, 72, 73, 74 illustrate an AAC encoder
in order to generate a lossy encoded bit stream referred to
as "perceptually coded bitstream" in Fig. 7. This bit
stream represents the base layer. In particular, block 71
in Fig. 7 designates the analysis filter bank including the

windowing with long and short windows according to the AAC
standard. Block 73 represents the quantization/encoding ac-
cording to the AAC standard and block 74 represents the bit
stream generation so that the bit stream on the output side
not only includes Huffman code words of quantized spectral
values but also the necessary side information, like for
example scale factors, etc., so that a decoding may be per-
formed. The lossy quantization in block 73 is here con-
trolled by the psychoacoustic model designated as the "per-
ceptual model" 72 in Fig. 7.
As already indicated, the output signal of block 74 is a
base scaling layer which requires relatively few bits and
is, however, only a lossy representation of the original
audio signal and may comprise encoder artifacts. The blocks
75, 76, 77, 78 represent the additional elements which are
needed to generate an extension bit stream which is loss-
less or virtually lossless, as it is indicated in Fig. 7.
In particular, the original audio signal is subjected to an
integer MDCT (IntMDCT) at the input 70, as it is illus-
trated by block 75. Further, the quantized spectral values,
generated by block 73, into which encoder losses are al-
ready introduced, are subjected to an inverse quantization
and to a subsequent rounding in order to obtain rounded
spectral values. Those are supplied to a difference former
77 forming a spectral-value-wise difference which is then
subjected to an entropy coding in block 78 in order to gen-
erate a lossless enhancement bit stream of the scaling
scheme in Fig. 7. A spectrum of differential values at the
output of block 77 thus represents the distortion intro-
duced by the psychoacoustic quantization in block 73.
On the decoder side the lossy coded bit stream or the per-
ceptually coded bit stream is supplied to a bit stream de-
coder 81. On the output side, block 81 provides a sequence
of blocks of quantized spectral values which are then sub-
jected to an inverse quantization in a block 82. At the
output of block 82 thus inversely quantized spectral values

are present which now, in contrast to the values at the in-
put of block 82, do not represent quantizer indices any-
more, but which are now so to say "correct" spectral values
which, however, are different from the spectral values be-
fore the encoding in block 73 of Fig. 7 due to the lossy
quantization. These quantized spectral values are now sup-
plied to a synthesis filter bank or an inverse MDCT trans-
formation (inverse MDCT), respectively, in block 83 to ob-
tain a psychoacoustically encoded and again decoded audio
signal (perceptual audio) which is different from the
original audio signal at the input 70 of Fig. 7 due to the
encoding errors introduced by the encoder of Fig. 7. In or-
der to not only obtain a lossy but even a lossless compres-
sion, the audio signal of block 82 is supplied to a round-
ing in a block 84. In an adder 85 now the rounded, in-
versely quantized spectral values are added to the differ-
ential values which were generated by the difference former
77, wherein in a block 86 an entropy decoding is performed
to decode the entropy code words contained in the extension
bit stream containing the lossless or virtually lossless
information, respectively.
At the output of block 85, IntMDCT spectral values are thus
present which are in the optimum case identical to the MDCT
spectral values at the output of block 75 of the encoder of
Fig. 7. The same are then subjected to an inverse integer
MDCT (inverse IntMDCT), to obtain a coded lossless audio
signal or virtually lossless audio signal (lossless audio)
at the output of block 87.
The integer MDCT (IntMDCT) is an approximation of the MDCT,
however, generating integer output values. It is derived
from the MDCT using the lifting scheme. This works in par-
ticular when the MDCT is divided into so-called Givens ro-
tations. Then, a two-stage algorithm with Givens rotations
and a subsequent DCT-IV result as the integer MDCT on the
encoder side and with a DCT-IV and a downstream number of
Givens rotations on the decoder side. In the scheme of Fig.

7 and Fig. 8, thus the quantized MDCT spectrum generated in
the AAC encoder is used to predicate the integer MDCT spec-
trum. In general, the integer MDCT is thus an example for
an integer transformation generating integer spectral val-
ues and again time samples from the integer spectral val-
ues, without losses being introduced by rounding errors.
Other integer transformations exist apart from the integer
MDCT.
The scaling scheme indicated in Figs. 7 and 8 is only suf-
ficiently efficient when the differences at the output of
the difference former 77 are small. In the scheme illus-
trated in Fig. 7 this is the case, as the MDCT and the in-
teger MDCT are similar and as the IntMDCT in block 75 is
derived from the MDCT in block 71, respectively. If this
was not the case, the scheme illustrated there would not be
suitable, as then the differential values would in many
cases be greater than the original MDCT values or even
greater than the original IntMDCT values. Then the scaling
scheme in Fig. 7 has lost its value as the extension scal-
ing layer output by block 78 has a high redundancy regard-
ing the base scaling layer.
Scalability schemes are always optimal when the base layer
comprises a number of bits and when the extension layer
comprises a number of bits and when the sum of the bits in
the base layer and in the extension layer is equal to a
number of bits which would be obtained if the base layer
already were a lossless encoding. This optimum case is
never achieved in practical scalability schemes, as for the
extension layer additional signaling bits are required.
This optimum is, however, aimed at as far as possible. As
the transformations in blocks 71 and 75 are relatively
similar in Fig. 7, the concept illustrated in Fig. 7 is
close to optimum.
This simple scalability concept may, however, not just like
that be applied to the output signal of an MP3 encoder, as

the MP3 encoder, as it was illustrated, comprises no pure
MDCT filter bank as a filter bank, but the hybrid filter
bank having a first filter bank stage for generating dif-
ferent subband signals and a downstream MDCT for further
breaking down the subband signals, wherein in addition, as
it is also indicated in the MP3 standard, an additional
aliasing cancellation stage of the hybrid filter bank is
implemented. As the integer MDCT in block 75 of Fig. 7 has
little similarities with the hybrid filter bank according
to the MP3 standard, a direct application of the concept
shown in Fig. 7 to an MP3 output signal would lead to very
high differential values at the output of the difference
former 77, which results in an extremely inefficient scal-
ability concept, as the extension layer requires far too
many bits in order to reasonably encode the differential
values at the output of the difference former 77.
A possibility for generating the extension bit stream for
an MP3 output signal is illustrated in Fig. 9 for the en-
coder and in Fig. 10 for the decoder. An MP3 encoder 90 en-
codes an audio signal and provides a base layer 91 on the
output side. The MP3 encoded audio signal is then supplied
to an MP3 decoder 92 providing a lossy audio signal in the
time range. This signal is then supplied to an IntMDCT
block which may in principle be setup just like block 75 in
Fig. 7, wherein this block 75 then provides IntMDCT spec-
tral values on the output side which are supplied to a dif-
ference former 77 which also includes IntMDCT spectral val-
ues as further input values, which were, however, not gen-
erated by the MP3 decoded audio signal but by the original
audio signal which was supplied to the MP3 encoder 90.
On the decoder side, the base layer is again supplied to an
MP3 decoder 92 to provide a lossy decoded audio signal at
an output 100 which would correspond to the signal at the
output of block 83 of Fig. 8. This signal would then have
to be subjected to an integer MDCT 75 to then be encoded
together with the extension layer 93 which was generated at

the output of the difference former 77. The lossless spec-
trum would then be present at an output 101 of the adder
102 and would only have to be converted by means of an in-
verse IntMDCT 103 into the time range in order to obtain a
losslessly decoded audio signal which would correspond to
the "lossless audio" at the beginning of block 87 of
Fig. 8.
The concept illustrated in Fig. 9 and in Fig. 10, which
provides a relatively efficiently encoded extension layer
just like the concepts illustrated in Figs. 7 and 8, is ex-
pensive both on the encoder side (Fig. 9) and also on the
decoder side (Fig. 10), respectively. In contrast to the
concept in Fig. 7, a complete MP3 decoder 92 and an addi-
tional IntMDCT 75 are required.
Another disadvantage in this scheme is, that a bit-accurate
MP3 decoder would have to be defined. This is not intended,
however, as the MP3 standard does not represent a bit-
accurate specification but only has to be fulfilled within
the scope of a "conformance" by a decoder.
On the decoder side, further a complete additional IntMDCT
stage 75 is required. Both additional elements cause compu-
tational overhead and are disadvantageous in particular for
use in mobile devices both with regard to chip consumption
and also current consumption and also with regard to the
associated delay.
In summary, advantages of the concept illustrated in Fig. 7
and Fig. 8 are, that compared to time domain methods no
complete decoding of the audio-adapted encoded signal is
required, and that an efficient encoding is obtained by a
representation of the quantization error in the frequency
range to be encoded additionally. Thus, the method stan-
dardized by ISO/IEC MPEG-4 Scalable Lossless Coding (SLS)
uses this approach, as described in R. Geiger, R. Yu, J.
Herre, S. Rahardja, S. Kim, X. Lin, M. Schmidt, "ISO/IEC

MPEG-4 High-Definition Scalable Advanced Audio Coding",
120th AES meeting, May 20 - 23, 2006, Paris, France, Pre-
print 6791. Thus, a backward compatible, lossless extension
of audio encoding methods, for example MPEG-2/4 AAC, is ob-
tained which use the MDCT as a filter bank.
This approach may, however, not directly be applied to the
widely used method MPEG-1/2 Layer 3 (MP3), as the hybrid
filter bank used in this method, in contrast to the MDCT,
is not compatible with the IntMDCT or another integer
transformation. Thus, a difference formation between the
decoded spectral values and the corresponding IntMDCT val-
ues in general does not lead to small differential values
and thus not to an efficient encoding of the differential
values. The core of the problem here is the time shifts be-
tween the corresponding modulation functions of the IntMDCT
and the MP3 hybrid filter bank. These lead to phase shifts
which in unfavorable cases even lead to the fact that the
differential values comprise higher values than the IntMDCT
values. Also an application of the principles underlying
the IntMDCT, like for example the lifting scheme, to the
hybrid filter bank of MP3 is problematic, as regarding its
basic approach - in contrast to MDCT - the hybrid filter
bank is a filter bank which provides no perfect reconstruc-
tion.
SUMMARY OF THE INVENTION
It is the object of the present invention to provide an ef-
ficient concept for processing audio data and in particular
for coding or decoding audio data.
This object is achieved by a device for postprocessing
spectral values according to claim 1, an encoder according
to claim 17, a decoder according to claim 22 or a method
according to claim 23, 24, 25 or a computer program accord-
ing to claim 26.

The present invention is based on the finding, that spec-
tral values, for example representing the base layer of a
scaling scheme, i.e. e.g. MP3 spectral values, are sub-
jected to postprocessing, to obtain values thereform which
are compatible with corresponding values obtained according
to an alternative transformation algorithm. According to
the invention, thus such a postprocessing is performed us-
ing weighted additions of spectral values so that the re-
sult of the postprocessing is as similar as possible to a
result which is obtained when the same audio signal is not
converted into a spectral representation using the first
transformation algorithm but using the second transforma-
tion algorithm, which is, in preferred embodiments of the
present invention, an integer transformation algorithm.
It is thus been found, that even with a strongly incompati-
ble first transformation algorithm and second transforma-
tion algorithm, by a weighted addition of certain spectral
values of the first transformation algorithm, a compatibil-
ity of the postprocessed values with the results of the
second transformation is achieved which is so good that an
efficient extension layer may be formed with differential
values, without the expensive and thus disadvantageous cod-
ing and decoding of the concept in Fig. 9 and Fig. 10 being
necessary. In particular, the weighted addition is per-
formed so that a postprocessed spectral value is generated
from a weighted addition of. a spectral value and an adja-
cent spectral value at the output of the first transforma-
tion algorithm, wherein preferably both spectral values
from adjacent frequency ranges and also spectral values
from adjacent time blocks or time periods, respectively,
are used. By the weighted addition of adjacent spectral
values it is considered that in the first transformation
algorithm adjacent filters of a filter bank overlap, as it
is the case virtually with all filter banks. By the use of
temporally adjacent spectral values, i.e. by the weighted
addition of spectral values (e.g. of the same or only a
slightly different frequency) of two subsequent blocks of

spectral values of the first transformation it is further
considered that typically transformation algorithms are
used in which a block overlap is used.
Preferably, the weighting factors are permanently pro-
grammed both on the encoder side and also on the decoder
side, so that no additional bits are necessary to transfer
weighting factors. Instead, the weighting factors are set
once and e.g. stored as a table or firmly implemented in
hardware, as the weighting factors are not signal-dependent
but only dependent on the first transformation algorithm
and on the second transformation algorithm. In particular,
it is preferred to set the weighting factors so that an im-
pulse response of the construction of first transformation
algorithm and postprocessing is equal to an impulse re-
sponse of the second transformation algorithm. In this re-
spect, an optimization of the weighting factors may be em-
ployed manually or computer-aided using known optimization
methods, for example using certain representative test sig-
nals or, as indicated, directly using the impulse responses
of the resulting filters.
The same postprocessing device may be used both on the en-
coder side and also on the decoder side in order to adapt
actually incompatible spectral values of the first trans-
formation algorithm to spectral values of the second trans-
formation algorithm, so that both blocks of spectral values
may be subjected to a difference formation in order to fi-
nally provide an extension layer for an audio signal which
is for example an MP3 encoded signal in the base layer and
comprises the lossless extension as the extension layer.
It is to be noted, that the present invention is not lim-
ited to the combination of MP3 and integer MDCT, but that
the present invention is of use everywhere, when spectral
values of actually incompatible transformation algorithms
are to be processed together, for example for the purpose
of a difference formation, an addition or any other combi-

nation operation in an audio encoder or an audio decoder.
The preferred use of the inventive postprocessing device
is, however, to provide an extension layer for a base layer
in which an audio signal is encoded with a certain quality,
wherein the extension layer, together with the base layer,
serves to achieve a higher-quality decoding, wherein this
higher-quality decoding preferably already is a lossless
decoding, but may, however, also be a virtually lossless
decoding, as long as the quality of the decoded audio sig-
nal is improved using the extension layer as compared to
the decoding using only the base layer.
BRIEF DESCRIPTION OF THE DRAWINGS
In the following, preferred embodiments of the present in-
vention are explained in more detail with reference to the
accompanying drawings, in which:
Fig. 1 shows an inventive device for postprocessing
spectral values;
Fig. 2 shows an encoder side of an inventive encoder
concept;
Fig. 3 shows a decoder side of an inventive decoder con-
cept;
Fig. 4 shows a detailed illustration of a preferred em-
bodiment of the inventive postprocessing and dif-
ference formation for long blocks;
Fig. 5a shows a preferred implementation of the inventive
postprocessing device for short blocks according
to a first variant;
Fig. 5b shows a schematical illustration of blocks of
values belonging together for the concept shown
in Fig. 5a;

Fig. 5c shows a sequence of windows for the variant shown
in Fig. 5a;
Fig. 6a shows a preferred implementation of the inventive
postprocessing device and difference formation
for short blocks according to a second variant of
the present invention;
Fig. 6b shows an illustration of diverse values for the
variant illustrated in Fig. 6a;
Fig. 6c shows a block raster for the variant illustrated
in Fig. 6a;
Fig. 7 shows a prior encoder illustration for generating
a scaled data stream;
Fig. 8 shows a prior decoder illustration for processing
a scaled data stream;
Fig. 9 shows an inefficient encoder variant; and
Fig. 10 shows an inefficient decoder variant.
DESCRIPTION OF THE PREFERRED EMBODIMENTS
Fig. 1 shows an inventive device for postprocessing spec-
tral values which are preferably a lossy representation of
an audio signal, wherein the spectral values have an under-
lying first transformation algorithm for converting the au-
dio signal into a spectral representation independent of
the fact whether they are lossy or not lossy. The inventive
device illustrated in Fig. 1 or the method also schemati-
cally illustrated in Fig. 1, respectively, distinguish
themselves - with reference to the device - by a means 12
for providing a sequence of blocks of spectral values rep-
resenting a sequence of blocks of samples of the audio sig-

nal. In a preferred embodiment of the present invention
which will be illustrated later, the sequence of blocks
provided by means 12 is a sequence of blocks generated by
an MP3 filter bank. The sequence of blocks of spectral val-
ues is supplied to an inventive combiner 13, wherein the
combiner is implemented to perform a weighted addition of
spectral values of the sequence of blocks of spectral val-
ues to obtain, on the output side, a sequence of blocks of
postprocessed spectral values, as it is illustrated by out-
put 14. In particular, the combiner 13 is implemented to
use, for calculating a postprocessed spectral value for a
frequency band and a time period, a spectral value of the
sequence of blocks for the frequency band and the time pe-
riod and a spectral value for an adjacent frequency band
and/or an adjacent time period. Further, the combiner is
implemented to use such weighting factors for weighting the
used spectral values, that the postprocessed spectral val-
ues are an approximation to spectral values obtained by a
second transformation algorithm for converting the audio
signal into a spectral representation, wherein, however,
the second transformation algorithm is different from the
first transformation algorithm.
This is illustrated by the schematical illustration in
Fig. 1 at the bottom. A first transformation algorithm is
represented by a reference numeral 16. The postprocessing,
as it is performed by the combiner, is represented by the
reference numeral 13, and the second transformation algo-
rithm is represented by a reference numeral 17. Of blocks
16, 13 and 17, blocks 16 and 17 are fixed and typically
mandatory due to external conditions. Only the weighting
factors of the postprocessing means 13 or the combiner 13,
respectively, represented by reference numeral 18, may be
set by the user. In this connection, this is not signal-
dependent but depending on the first transformation algo-
rithm and the second transformation algorithm, however. By
the weighting factors 18 it may further be set, how many
spectral values adjacent regarding frequency or spectral

values adjacent in time are combined with each other. If a
weighting factor, as it will be explained with reference to
Figs. 4 to 6, is set to 0, the spectral value associated
with this weighting factor is not considered in the combi-
nation .
In preferred embodiments of the present invention, for each
spectral value a set of weighting factors is provided.
Thus, a considerable amount of weighting factors result.
This is unproblematic, however, as the weighting factors do
not have to be transferred but only have to be permanently
programmed to the encoder side and the decoder side. If en-
coder and decoder thus agreed on the same set of weighting
factors for each spectral value and, if applicable, for
each time period, or, as it will be illustrated in the fol-
lowing, for each subblock or ordering position, respec-
tively, no signaling has to be used for the present inven-
tion, so that the inventive concept achieves a substantial
reduction of the data rate in the extension layer without
any signaling of additional information, without any accom-
panying quality losses.
The present invention thus provides a compensation of the
phase shifts between frequency values, as they are obtained
by the first transformation algorithm, and frequency val-
ues, as they are obtained by the second transformation al-
gorithm, wherein this compensation of the phase shifts may
be presented via a complex spectral representation. For
this purpose, the concept described in DE 10234130 is in-
cluded for reasons of clarity, in which for calculating
imaginary parts from real filter bank output values linear
combinations of temporally and spectrally adjacent spectral
values are obtained. If this procedure was used for decoded
MP3 spectral values, a complex-valued spectral representa-
tion would be obtained. Each of the resulting complex spec-
tral values may now be modified in its phase position by a
multiplication by a complex-valued correction factor so
that, according to the present invention, it gets as close

to the second transformation algorithm as possible, i.e.
preferably the corresponding IntMDCT value, and is thus
suitable for a difference formation. Further, according to
the invention, also a possibly required amplitude correc-
tion is performed. According to the invention, these steps
for the formation of the complex-valued spectral represen-
tation and the phase or sum correction, respectively, are
summarized such that by the linear combination of spectral
values on the basis of the first transformation algorithm
and its temporal and spectral neighbors a new spectral
value is formed which minimizes the difference to the cor-
responding IntMDCT value. According to the invention, in
contrast to the DE 10234130, a postprocessing of filter
bank output values is not performed using weighting factors
in order to obtain real and imaginary parts. Instead, ac-
cording to the invention a postprocessing is performed us-
ing such weighting factdrs that, as it was illustrated in
Fig. 1 at the bottom, a combination of the first transfor-
mation algorithm 16 and the postprocessing 13 is set by the
weighting factors so that the result corresponds to a sec-
ond transformation algorithm as far as possible.
Fig. 2 and Fig. 3 show a preferred field of use of the in-
ventive concept illustrated in Fig. 1 both on the encoder
side (Fig. 2) and also on the decoder side (Fig. 3) of a
scalable encoder. An MP3 bit stream 20 or - generally - a
bit stream, respectively, as it may be obtained by a first
transformation algorithm, is fed to a block 21 in order to
generate the spectral values from the bit stream which are
for example MP3 spectral values. The decoding of the spec-
tral values in block 21 will thus typically include an en-
tropy decoding and an inverse quantization.
Then, in block 10, a calculation of approximation values is
performed, wherein the calculation of approximation values
or of blocks of postprocessed spectral values, respec-
tively, is performed like it was illustrated in Fig. 1.
Hereupon, a difference formation is performed in a block

22, using IntMDCT spectral values, as they are obtained by
an IntMDCT conversion in a block 23. Block 23 thus obtains
an audio signal as an input signal from which the MP3 bit
stream, like it is fed into the input 20, was obtained by
encoding. Preferably, the differential spectrums as they
are obtained by block 22 are subjected to a lossless encod-
ing 24 which for example includes a delta encoding, a Huff-
man encoding, an arithmetic encoding or any other entropy
coding by which the data rate is reduced, no losses are in-
troduced into a signal, however.
On the decoder side, the MP3 bit stream 20, as it was also
fed into the input 20 of Fig. 2, is again subjected to a
decoding of the spectral values by a block 21, which may
correspond to block 21 of Fig. 2. Hereupon, the MP3 spec-
tral values obtained at the output of block 21 are again
processed according to Fig. 1 or block 10. On the decoder
side, however, the blocks of postprocessed spectral values,
as they are output by block 10, are supplied to an addition
stage 30, which obtains IntMDCT differential values at its
other input, as they are obtained by a lossless decoding 31
from the lossless extension bit stream which was output by
block 24 in Fig. 2. By the addition of the IntMDCT differ-
ential values output by block 31 and the processed spectral
values output by block 10, then, at an output 32 of the ad-
dition stage 30 blocks of IntMDCT spectral values are ob-
tained which are a lossless representation of the original
audio signal, i.e. of the audio signal which was input into
block 23 of Fig. 2. The lossless audio output signal is
then generated by a block 33 which performs an inverse
IntMDCT in order to obtain a lossless or virtually lossless
audio output signal. Generally speaking, the audio output
signal at the output of block 33 has a better quality than
the audio signal which would be obtained if the output sig-
nal of block 21 was processed with an MP3 synthesis hybrid
filter bank. Depending on the implementation, the audio
output signal at output 33 may thus be an identical repro-
duction of the audio signal which was input into block 23

of Fig. 2, or a representation of this audio signal, which
is not identical, i.e. not completely lossless, which has,
however, already a better quality than a normal MP3 coded
audio signal.
At this point it is to be noted, that as a first transfor-
mation algorithm the MP3 transformation algorithm with its
hybrid filter bank is preferred, and that as a second
transformation algorithm the IntMDCT algorithm as an inte-
ger transformation algorithm is preferred. The present in-
vention is already advantageous everywhere, however, where
two transformation algorithms are different from each
other, wherein both transformation algorithms do not neces-
sarily have to be integer transformation algorithms within
the scope of the IntMDCT transformation, but may also be
normal transformation algorithms which are, within the
scope of an MDCT, not necessarily an invertible integer
transformation. According to the invention it is preferred,
however, that the first transformation algorithm is a non-
integer transformation algorithm and that the second trans-
formation algorithm is an integer transformation algorithm,
wherein the inventive postprocessing is in particular ad-
vantageous when the first transformation algorithm provides
spectrums which are, compared to the spectrums provided by
the second transformation algorithm, phase shifted and/or
changed with regard to their amounts. In particular when
the first transformation algorithm is not even perfectly
reconstructing, the inventive simple postprocessing by a
linear combination is especially advantageous and may effi-
ciently be used.
Fig. 4 shows a preferred implementation of the combiner 13
within an encoder. The implementation within a decoder is
identical, however, if the adder 22 does not, like in
Fig. 4, perform a difference formation, as it is illus-
trated by the minus sign above the adder 22, but when an
addition operation is performed, as it is illustrated in
block 30 of Fig. 3. In each case the values which are fed

into an input 40 are values as they are obtained by the
second transformation algorithm 23 of Fig. 2 for the en-
coder implementation or as they are obtained by block 31 of
Fig. 3 in the decoder implementation.
In a preferred embodiment of the present invention, the
combiner includes three sections 41, 42, 43. Each section
includes three multipliers 42a, 42b, 42c, wherein each mul-
tiplier is associated with a spectral value with a fre-
quency index k-1, k or k+1. Thus, the multiplier 42a is as-
sociated with the frequency index k-1. The multiplier 42b
is associated with the frequency index k and the multiplier
42c is associated with the frequency index k+1.
Each branch thus serves for weighting spectral values of a
current block with the block index v or n+1, n or n-1, re-
spectively, in order to obtain weighted spectral values for
the current block.
Thus, the second section 42 serves for weighting spectral
values of a temporally preceding block or temporally subse-
quent block. With regard to section 41, section 42 serves
for weighting spectral values of the block n temporally
following block n+1, and section 43 serves for weighting
the block n-1 following block n. In order to indicate this,
delay elements 44 are indicated in Fig. 4. For reasons of
clarity, only one delay element "z"1" is designated by the
reference numeral 44.
In particular, each multiplier is provided with a spectral
index-dependent weighting factor co(k) to c8(k). Thus, in
the preferred embodiment of the present invention, nine
weighted spectral values result, from which a postprocessed
spectral value y is calculated for the frequency index k
and the time block n. These nine weighted spectral values
are summed up in a block 45.

The postprocessed spectral value for the frequency index k
and the time index n is thus calculated by the addition of
possibly differently weighted spectral values of the tempo-
rally preceding block (n-1) and the temporally subsequent
block (n+1) and using respectively upwardly (k+1) and down-
wardly (k-1) adjacent spectral values. More simple imple-
mentations may only be, however, that a spectral value for
the frequency index k is combined only with one adjacent
spectral value k+1 or k-1 from the same block, wherein this
spectral value which is combined with the spectral value of
the frequency index k, does not necessarily have to be di-
rectly adjacent but may also be a different spectral value
from the block. Due to the typical overlap of adjacent
bands it is preferred, however, to perform a combination
with the directly adjacent spectral value to the top and/or
to the bottom.
Further, alternatively or additionally, each spectral value
with a spectral value for a different time duration, i.e. a
different block index, may be combined with the correspond-
ing spectral value from block n, wherein this spectral
value from a different block does not necessarily have to
have the same frequency index but may have a different,
e.g. adjacent frequency index. Preferably, however, at
least the spectral value with the same frequency index from
a different block is combined with the spectral value from
the currently regarded block. This other block again does
not necessarily have to be the direct temporally adjacent
one, although this is especially preferable when the first
transformation algorithm and/or the second transformation
algorithm have a block overlap characteristic, as it is
typical for MP3 encoders or AAC encoders.
This means, when the weighting factors of Fig. 4 are con-
sidered, that at least the weighting factor c4(k) is un-
equal 0, and that at least a second weighting factor is un-
equal 0, while all other weighting factors may also be
equal to 0, which may also already provide a processing,

which may, however, due to the low number of weighting fac-
tors unequal 0 only be a relatively coarse approximation of
the second transformation algorithm, if again the bottom
half of Fig. 1 is regarded. In order to consider more than
nine spectral values, further branches for blocks further
in the future or further in the past may be added. Further,
also further multipliers and further corresponding weight-
ing factors for spectral values lying spectrally farther
apart may be added, to generate a field from the 3x3 field
of Fig. 4, which comprises more than three lines and/or
more than three columns. It has been found, however, that
when nine weighting factors are admitted for each spectral
value, compared to a lower number of weighting factors,
substantial improvements are achieved, while when the num-
ber of weighting factors is increased, no substantial fur-
ther improvements regarding decreasing differential values
at the outputs of block 22 are obtained, so that a greater
number of weighting factors with typical transformation al-
gorithms with an overlap of adjacent subband filters and a
temporary overlap of adjacent blocks brings no substantial
improvements.
Regarding the 50 percent overlap used in the sequence of
long blocks, reference is made to the schematical illustra-
tion of Fig. 5c at 45 at the left of the figure, where two
subsequent long blocks are illustrated schematically. The
combiner concept illustrated in Fig. 4 is thus always used,
according to the invention, when a sequence of long blocks
is used, wherein the block length of the IntMDCT algorithm
23 and the degree of overlap of the IntMDCT algorithm is
set equal to the degree of overlap of the MP3 analysis fil-
ter and the block length of the MP3 analysis filter. In
general it is preferred that block overlap and block length
of both transformation algorithms are set equally, which
presents no special limitation, as the second transforma-
tion algorithm, i.e. for example the IntMDCT 23 of Fig. 2,
may easily be set with regard to those parameters, while
the same is not easily possible with the first transforma-

tion algorithm, in particular when the first transformation
algorithm is standardized as with regard to the example of
MP3 and is frequently used and may thus not be changed.
As it was already illustrated with reference to Fig. 2 and
Fig. 3, the associated decoder in Fig. 3 reverses the dif-
ference formation again by an addition of the same approxi-
mation values, i.e. the IntMDCT differential values at the
output of block 22 of Fig. 2 or at the output of block 31
of Fig. 3.
According to the invention, this method may thus generally
be applied to the difference formation between spectral
representations obtained using different filter banks, i.e.
when one filter bank/transformation underlying the first
transformation algorithm is different from a filter
bank/transformation underlying the second transformation
algorithm.
One example for the concrete application is the use of the
MP3 spectral values from "long block" in connection with an
IntMDCT, as it was described with reference to Fig. 4. As
the frequency resolution of the hybrid filter bank in this
case is 576, the IntMDCT will also comprise a frequency
resolution of 576, so that the window length may comprise a
maximum of 1152 time samples.
In the example described in the following, only the direct
temporal and spectral neighbors are used, while in the gen-
eral case also (or alternatively) values being farther
apart may be used.
If the spectral value of the k-th band in the n-th MP3
block is designated by x(k,n) and the corresponding spec-
tral value of the IntMDCT is designated by y(k,n), the dif-
ference is calculated as illustrated in Fig. 4 for d(k,n).
y(k,n) is the approximation value for y(k,n) obtained by

the linear combination, and is determined as it is illus-
trated by the long equation below Fig. 4.
It is to be noted here, that due to the different phase
difference for each of the 576 subbands a distinct coeffi-
cient set may be required. In the practical realization, as
it is illustrated in Fig. 4, for an access to temporally
adjacent spectral values delays 44 are used whose output
values respectively correspond to input values in a corre-
sponding preceding block. In order to enable an access to
temporally subsequent spectral values, thus also the
IntMDCT spectral values as they are applied to the input 40
are delayed by a delay 46.
Fig. 5a shows a somewhat modified procedure when the MP3
hybrid filter bank provides short blocks wherein three sub-
blocks respectively are generated by 192 spectral values,
wherein here apart from the first variant of Fig. 5a also a
second variant in Fig. 6a is preferred according to the in-
vention.
The first variant is based on a triple application of an
IntMDCT with a frequency resolution 192 for forming corre-
sponding blocks of spectral values. Here, the approximation
values may be formed from the three values belonging to a
frequency index and their corresponding spectral neighbors.
For each subblock, here a distinct set of coefficients is
required. For describing the procedure thus a subblock in-
dex u is introduced, so that n again corresponds to the in-
dex of a complete block of the length 576. Expressed as an
equation, thus the system of equations of Fig. 5a results.
Such a sequence of blocks is illustrated in Fig. 5b with
reference to the values and in Fig. 5c with reference to
the windows. The MP3 encoder provides short MP3 blocks, as
they are illustrated at 50. The first variant also provides
short IntMDCT blocks y(u0), y(ux) and y(u2), as it is illus-
trated at 51 in Fig. 5b. By this, three short differential
blocks 52 may be calculated such that a 1:1 representation

results between a corresponding spectral value at the fre-
quency k in blocks 50, 51 and 52.
In contrast to Fig. 4 it is to be noted, that in Fig. 5a
delays 44 are not indicated. This results from the fact
that the postprocessing may only be performed when all
three subblocks 0, 1, 2 for a block n have been calculated.
If the subblock with the index 0 is the temporally first
subblock, and if the next subblock with the index 1 is the
temporally later block, and if the index u=2 is the again
temporally later short block, then the differential block
for the index u=0 is calculated using spectral values from
the subblock u0, the subblock Ui and the subblock u2. This
means, that only with reference to the currently calculated
subblock with the index 0 future subblocks 1 and 2 are
used, however no spectral value from the past. This is sen-
sible, as a switch to short blocks was performed, as there
was a transient result in the audio signal as it is known
and for example illustrated in the above-mentioned expert's
publication of Edler. The postprocessed values for the sub-
block having the index 1 used for gaining the differential
values having the subblock index 1 are, however, calculated
from a temporally preceding, from a temporally current and
from a temporally subsequent subblock, while the postproc-
essed spectral values for the third subblock with the index
2 are not calculated using future subblocks but only using
past subblocks having the index 1 and the index 0, which is
also technically sensible in so far as again, as indicated
in Fig. 5c, easily a window switch to long windows may be
initiated by a stop window, so that later again a change
directly to the long block scheme of Fig. 4 may be per-
formed.
Fig. 5 makes thus clear that in particular with short
blocks, however also generally, it may be sensible to look
only into the past or into the future and not always, as
indicated in Fig. 4, both into the past and also into the

future, to obtain spectral values which provide a postproc-
essed spectral value after a weighting and a summation.
In the following, with reference to Fig. 6a, 6b and 6c the
second variant for short blocks is illustrated. In the sec-
ond variant, the frequency resolution of the IntMDCT is
still 576, so that three spectrally adjacent IntMDCT spec-
tral values each lie in the frequency range of one MP3
spectral value. Thus, for each of those three IntMDCT spec-
tral values, for a difference formation a distinct linear
combination is formed from the three temporally subsequent
subblock spectral values and their spectral neighbors,
wherein the index s which is also referred to as an order
index now indicates the position within • each group of
three. Thus, the equation as it is illustrated in Fig. 6a
below the block diagram results. This second variant is es-
pecially suitable if a window function with a small overlap
area is used in the IntMDCT, as then the considered signal
section corresponds well to that of the three subblocks. In
this case, like with the first variant, it is preferred to
adapt the window forms of the IntMDCT of preceding or sub-
sequent long blocks, respectively, so that a perfect recon-
struction results. A corresponding block diagram for the
first variant is illustrated in Fig. 5c. A corresponding
diagram for the second variant is illustrated in Fig. 6c,
wherein now only one single long IntMDCT block is generated
by the long window 63, wherein this long IntMDCT block now
comprises k triple blocks of spectral values, wherein the
bandwidth of such a triple block resulting from s=0, s=l
and s=2 is equal to the bandwidth of a block k of the short
MP3 blocks 60 in Fig. 6b. From Fig. 6a it may be seen that
for a subtraction from the first spectral value with s=0
for a triple block having the index k again the values of
the current, the future and the next future subblock 0, 1,
2 are used, however, no values from the past are used. For
calculating a differential value for the second value s=l
of a triple group, however, spectral values from the pre-
ceding subblock and the future subblock are used, while for

calculating a differential spectral value having the order
index s = 2 only preceding subblocks are used, as it is il-
lustrated by branches 41 and 42 which are in the past with
reference to branch 43 in Fig. 6a.
At this point it is to be noted that with all calculation
regulations the terms exceeding the limits of the frequency
range, i.e. e.g. the frequency index -1 or 576 or 192, re-
spectively, are each omitted. In these cases, in the gen-
eral example in Figs. 4 to 6 the linear combination is thus
reduced to 6 instead of 9 terms.
In the following, detailed reference is made to the window
sequences in Fig. 5c and Fig. 6c. The window sequences con-
sist of a sequence of long blocks, as they are processed by
the scenario in Fig. 4. Hereupon, a start window 56 follows
having an asymmetrical form, as it is "converted" from a
long overlapping area at the beginning of the start window
to a short overlapping area at the end of the start window.
Analog to this, a stop window 57 exists which is again con-
verted from a sequence of short blocks to a sequence of
long blocks and thus comprises a short overlapping area at
the beginning and a long overlapping area at the end.
A window switch is, as it is illustrated in the mentioned
expert's publication of Edler, selected if a time duration
in the audio .signal is detected by an encoder which com-
prises a transient signal.
Such a signaling is located in the MP3 bit stream, so that
when the IntMDCT, according to Fig. 2 and according to the
first variant.of Fig. 5c, also switches to short blocks, no
distinct transient detection is necessary, but a transient
detection based only on a short window notice in the MP3
bit stream takes place. For the postprocessing of values in
the start window it is preferred, due to the long overlap-
ping area with the preceding window, to use blocks with the
preceding block index n-1, while blocks with the subsequent

block index are only lightly weighted or generally not used
due to the short overlapping area. Analog to this, the stop
window for postprocessing will only consider values with a
future block index n+1 in addition to the values for the
current block n, but will only perform a weak weighting or
a weighting equal to 0, i.e. no use from the past, i.e.
e.g. from the third short block.
When, as shown in Fig. 6c, the sequence of windows as it is
implemented by the IntMDCT 23, i.e. the second transforma-
tion algorithm, performs no switch to short windows, how-
ever implements the preferably used window switch, then it
is preferred to initiate or terminate, respectively, the
window with the short overlap, designated by 63 in Fig. 6c,
also by a start window 56 and by a stop window 57.
Although in the embodiment illustrated in Fig. 6c the
IntMDCT of Fig. 2 does not change into the short window
mode, the signaling of short windows in the MP3 bit stream
may anyway be used to activate the window switch with a
start window, window with short overlap, as it is indicated
in Fig. 6c at 63, and stop window.
Further it is to be noted, that in particular the window
sequences illustrated in the AAC standard, adapted to the
MP3 block length or the MP3 feed, respectively, of 576 val-
ues for long blocks and 192 values for short blocks, and in
particular also the start windows and stop windows illus-
trated there, are especially suitable for an implementation
of the IntMDCT in block 23 of the present invention.
In the following, reference is made to the accuracy of the
approximation of first transformation algorithm and post-
processing.
For 576 input signals respectively having one impulse at
the position 0 ... 575 within a block, the following steps
were performed:

calculating the hybrid filter bank + approximation
- calculating the MDCT
calculating the square sum of the MDCT spectral compo-
nents
calculating the square sum of the deviations between
MDCT spectral components and the approximation. Here,
the maximum square deviation across all 576 signals is
determined.
The maximum relative square deviation across all positions
was, when using
long blocks according to Fig. 4, approx. 3.3 %
short blocks (hybrid) and long blocks (MDCT) according
to Fig. 6, approx. 20.6 %.
One could thus say, that with an impulse at the inputs of
the two transformations, the square sum of the deviations
between the approximation and the spectral components of
the second transformation should not be more than 30% (and
preferably not even more than 25% or 10% respectively) of
the square sum of the spectral components of the second
transformation, independent of the position of the impulse
in the input block. For calculating the square sums, all
blocks of spectral components should be considered which
are influenced by the impulse.
It is to be noted, that in the above error inspection (MDCT
versus hybrid FB + postprocessing) always the relative er-
ror was considered which is signal independent.
In the IntMDCT (versus MDCT), however, the absolute error
is signal independent and lies in a range of around -2 to 2
of the rounded integer values. From this it results that
the relative error becomes signal dependent. In order to
eliminate this signal dependency, preferably a fully con-

trolled impulse is assumed (e.g. value 32767 at 16 bit
PCM) .
This will then result in a virtually flat spectrum with an
average amplitude of about 32767/sqrt (576) = 1365 (energy
conservation) . The mean square error would then be about
2A2/1365/N2 = 0.0002%, i.e. negligible.
With a very low impulse at the input, the error would be
drastical, however. An impulse of the amplitude 1 or 2
would virtually completely be lost in the IntMDCT approxi-
mation error.
The error criterion of the accuracy of the approximation,
i.e. the value desired for the weighting factors, is thus
best comparable, when it is indicated for a fully con-
trolled impulse.
Depending on the circumstances, the inventive method may be
implemented in hardware or in software. The implementation
may take place on a digital storage medium, in particular a
floppy disc or a CD having electronically readable control
signals, which may cooperate with a programmable computer
system so that the method is performed. In general, the in-
vention thus also consists in a computer program product
having a program code stored on a machine-readable carrier
for performing the inventive method, when the computer pro-
gram product runs on a computer. In other words, the inven-
tion may thus be realized as a computer program having a
program code for performing the method, when the computer
program runs on a computer.

Claims
A device (10) for postprocessing spectral values based
on a first transformation algorithm (16) for convert-
ing an audio signal into a spectral representation,
comprising:
a means (12) for providing a sequence of blocks of the
spectral values representing a sequence of blocks of
samples of the audio signal; and
a combiner (13) for weightedly adding spectral values
of the sequence of blocks of spectral values in order
to obtain a sequence of blocks of postprocessed spec-
tral values, wherein the combiner (13) is implemented
to use, for the calculation of a postprocessed spec-
tral value for a frequency band and a time duration, a
spectral value of the sequence of blocks for the fre-
quency band and the time duration, and a spectral
value for another frequency band or another time dura-
tion, and wherein the combiner (13) is implemented to
use such weighting factors when weightedly adding,
that the postprocessed spectral values are an approxi-
mation to spectral values as they are obtained by a
second transformation algorithm (17) for converting
the audio signal into a spectral representation,
wherein the second transformation algorithm (17) is
different from the first transformation algorithm
(16) .
The device according to claim 1, wherein the first
transformation algorithm (16) is a hybrid transforma-
tion algorithm having two stages, and the second
transformation algorithm (17) is a one-stage transfor-
mation algorithm.
The device according to claim 1, wherein the first
transformation algorithm (16) comprises a polyphase

filter bank and a modified discrete cosine transforma-
tion, and wherein the second transformation algorithm
(17) is an integer MDCT.
4. The device according to claim 1, wherein the first
transformation algorithm (16) and the second transfor-
mation algorithm (17) are implemented so that they
provide real output signals.
5. The device according to claim 1, wherein the combiner
(13) is implemented to use such weighting factors that
the first transformation algorithm (16) and a post-
processing performed by the combiner (13) together
provide an impulse response which approximates an im-
pulse response of the second transformation algorithm
(17) .
6. The device according to claim 5, wherein in an ap-
proximation from the first transformation algorithm
and postprocessing, the weighting factors are selected
such that with an impulse at the input of the two
transformations the square sum of the deviations be-
tween the approximation and the spectral components of
the second transformation is no more than 30% of the
square sum of the spectral components of the second
transformation.
7. The device according to claim 1, wherein the means
(12) for providing a sequence of blocks is implemented
to provide blocks which are a lossy representation of
the audio signal.
8. The device according to claim 1, wherein the combiner
(13) for a calculation of a postprocessed spectral
value for a frequency band k comprises:
a first section (41, 42, 43) for weighting spectral
values of a current block for the frequency band k, a

frequency band k-1 or a frequency band k+1, in order
to obtain weighted spectral values for the current
block;
a second section (41, 42, 43) for weighting spectral
values of a temporally preceding block k-1 or tempo-
rally subsequent block k+1, in order to obtain
weighted spectral values for the temporally preceding
block or the temporally subsequent block; and
a means (45) for adding the weighted spectral values
to obtain a postprocessed spectral value for the fre-
quency band k of a current or preceding or subsequent
block of postprocessed spectral values.
9. The device according to claim 8, further comprising:
a third section (43) for weighting spectral values of
a preceding block, wherein the first section is imple-
mented to weight spectral values of a subsequent
block, and wherein the second section (42) is imple-
mented to weight spectral values of a current block,
and wherein the summer (45) is implemented to add
weighted spectral values of the three sections in or-
der to obtain a postprocessed spectral value for the
current block of postprocessed spectral values.
10. The device according to claim 1,
wherein the first transformation algorithm comprises a
block overlap function, wherein blocks of samples of
the time audio signal which the sequence of blocks of
spectral values is based on overlap.
11. The device according to claim 1, wherein the combiner
(13) is implemented to use a signal independent set of
weighting factors for each spectral value.

12. The device according to claim 1, wherein the sequence
of blocks of the spectral values comprises a set of
blocks of spectral values which are shorter than a
long block of spectral values which follows after the
set of blocks or which precedes the set of blocks, and
wherein the combiner (13) is implemented to use the
same frequency band or an adjacent frequency band out
of several blocks of the set of short blocks for cal-
culating a postprocessed spectral value for the set of
blocks of spectral values.
13. The device according to claim 12, wherein the combiner
(13) is implemented to use only spectral values of
short blocks and no spectral value of a preceding long
block or a subsequent long block for calculating post-
processed spectral values due to short blocks of spec-
tral values.
14. The device according to claim 1, wherein the combiner
(13) is implemented to implement the following equa-
tion:

wherein ŷ(k,n) is a postprocessed spectral value for
a frequency index k and a time index n, wherein x(k,n)
is a spectral value of a block of spectral values with
a frequency index k and a time index n, wherein co(k),
..., c8(k) are weighting factors, associated with the
frequency index k, wherein k-1 is a decremented fre-
quency index, wherein k+1 is an incremented frequency
index, wherein n-1 is a decremented time index and
wherein n+1 is an incremented time index.

The device according to claim 1, wherein the combiner
(13) is implemented to implement the following equa-
tion:

wherein ŷ (k,n,u) is a postprocessed spectral value
for a frequency index k and a time index n and a sub-
block index u, wherein x(k,n,u) is a spectral value of
a block of spectral values with a frequency index k
and a time index n and a subblock index u, wherein
c0(k), ..., c8(k) are weighting factors associated with
the frequency index k, wherein k-1 is a decremented
frequency index, wherein k+1 is an incremented fre-
quency index, wherein n-1 is a decremented time index
and wherein n+1 is an incremented time index, wherein
u is a subblock index indicating a position of a sub-
block in a sequence of subblocks, and wherein the time
index specifies a long block and the subblock index
specifies a comparatively short block.
The device according to claim 1, wherein the combiner
(13) is implemented in order to implement the follow-
ing equation:

wherein ŷ(k,n) is a postprocessed spectral value for
a frequency index k and a time index n, wherein
x(k,n,u) is a spectral value of a block of spectral
values with a frequency index k and a time index n and
a subblock index u, wherein c0(k), ..., c8(k) are

weighting factors associated with the frequency index
k, wherein k-1 is a decremented frequency index,
wherein k+1 is an incremented frequency index, wherein
n-1 is a decremented time index and wherein n+1 is an
incremented time index, wherein s is a order index in-
dicating a position of a subblock in a sequence of
subblocks, and wherein the time index specifies a long
block and the subblock index specifies a comparatively
short block.
17. An encoder for encoding an audio signal, comprising:
a device (10) for postprocessing spectral values ac-
cording to claim 1;
a means (23) for calculating a sequence of blocks of
spectral values according to the second transformation
algorithm (17) from the audio signal;
a means (22) for a spectral-value-wise difference for-
mation between the sequence of blocks due to the sec-
ond transformation algorithm and the sequence of
blocks of postprocessed spectral values.
18. The encoder according to claim 17, further comprising:
a means (24) for generating an extension bit stream
due to a result generated by the means (22) for a
spectral-value-wise difference formation.
19. The encoder according to claim 18, wherein the means
(24) for generating comprises an entropy encoder.
20. The encoder according to claim 17, wherein the se-
quence of blocks due to the first transformation algo-
rithm (16) is based on a lossy compression, and
wherein the sequence of blocks due to the second

transformation algorithm (17) is based on a lossless
or virtually lossless compression.
21. The encoder according to claim 17, comprising a memory
for storing the weighting factors in which the weight-
ing factors are storable independent of a signal.
22. The encoder according to claim 17, wherein the means
(23) for generating the sequence of blocks using the
second transformation algorithm (17) is implemented to
perform a windowing with a window sequence which de-
pends on a window sequence which the sequence of
blocks of the spectral values is based on which is
given due to the first transformation algorithm (16).
23. The encoder according to claim 22, wherein the means
(23) for providing a sequence of blocks using the sec-
ond transformation algorithm (17) is implemented to
switch from a long window with a long overlapping area
to a long window with a short overlapping area or to a
plurality of short windows, when in the sequence of
blocks of the spectral values due to the first trans-
formation algorithm (16) a switch to short windows
takes place.
24. A decoder for decoding an encoded audio signal, com-
prising:
a device for postprocessing spectral values according
to claim 1;
a means (31) for providing spectral-value-wise differ-
ential values between a sequence of blocks of post-
processed spectral values due to the first transforma-
tion algorithm (16) and a sequence of blocks due to
the second transformation algorithm (17);

a means (30) for combining the sequence of blocks of
the postprocessed spectral values and the differential
values in order to obtain a sequence of blocks of com-
bination spectral values; and
a means (33) for inversely transforming the sequence
of blocks of combination spectral values according to
the second transformation algorithm (17) to obtain a
decoded audio signal.
25. A method (10) for postprocessing spectral values which
are based on a first transformation algorithm (16) for
converting an audio signal into a spectral representa-
tion, comprising the following steps:
providing (12) a sequence of blocks of the spectral
values representing a sequence of blocks of samples of
the audio signal; and
weightedly adding (13) of spectral values of the se-
quence of blocks of spectral values to obtain a se-
quence of blocks of postprocessed spectral values,
wherein for calculating a postprocessed spectral value
for a frequency band and a time duration a spectral
value of the sequence of blocks for the frequency band
and the time duration and a spectral value for another
frequency band or another time duration are used, and
wherein such weighting factors are used when weight-
edly adding so that the postprocessed spectral values
are an approximation to spectral values as they are
obtained by a second transformation algorithm (17) for
converting the audio signal into a spectral represen-
tation, wherein the second transformation algorithm
(17) is different from the first transformation algo-
rithm (16) .
26. A method for encoding an audio signal, comprising the
following steps:

postprocessing (10) of spectral values according to
claim 25;
calculating (23) a sequence of blocks of spectral val-
ues according to the second transformation algorithm
(17) from the audio signal;
spectral-value-wise difference formation (22) between
the sequence of blocks of spectral values due to the
second transformation algorithm and the sequence of
blocks of postprocessed spectral values.
27. A method for decoding an encoded audio signal, com-
prising the following steps:
postprocessing spectral values according to claim 25;
providing (31) of spectral-value-wise differential
values between a sequence of blocks of postprocessed
spectral values due to the first transformation algo-
rithm (16) and a sequence of blocks of spectral values
due to the second transformation algorithm (17);
combining (30) the sequence of blocks of the postproc-
essed spectral values and the differential values to
obtain a sequence of blocks of combination spectral
values; and
inversely transforming (33) the sequence of blocks of
combination spectral values according to the second
transformation algorithm (17) to obtain a decoded au-
dio signal.
28. A computer program having a program code for perform-
ing the method of claim 25, when the computer program
runs on a computer.

29. A bit stream extension layer for inputting into an au-
dio decoder, wherein the bit stream extension layer
comprises a sequence of blocks of differential values,
wherein a block of differential values comprises,
spectral-value-wise, a difference between a block of
spectral values as it is obtained from a second trans-
formation algorithm (17) and a block of postprocessed
spectral values, wherein the postprocessed spectral
values are generated by a weighted adding of spectral
values of a sequence of blocks, as they are obtained
from a first transformation algorithm (16), wherein
for calculating a postprocessed spectral value for a
frequency band and a time duration, a spectral value
of the sequence of blocks for the frequency band and
the time duration and a spectral value for another
frequency band or another time duration are used, and
wherein for combining weighting factors are used such
that the postprocessed spectral values represent an
approximation to spectral values as they are obtained
by the second transformation algorithm (17), wherein
the second transformation algorithm (17) is different
from the first transformation algorithm (16).

For postprocessing spectral values which are based on a
first transformation algorithm for converting the audio
signal into a spectral representation, first a sequence of
blocks of the spectral values representing a sequence of
blocks of samples of the audio signal are provided. Here-
upon, a weighted addition of spectral values of the se-
quence of blocks of spectral values is performed in order
to obtain a sequence of blocks of postprocessed spectral
values, wherein the combination is performed such that for
calculating a postprocessed spectral value for a frequency
band and a time duration a spectral value of the sequence
of blocks for the frequency band and the time duration and
a spectral value for another frequency band or another time
duration are used, wherein the combination is further per-
formed such that such weighting factors are used that the
postprocessed spectral values are an approximation to the
spectral values as they are obtained by converting the au-
dio signal into a spectral representation using a second
transformation algorithm which is different from the first
transformation algorithm. The postprocessed spectral values
are in particular used for a difference formation within a
scalable encoder or for an addition within a scalable decoder, respectively.

Documents:

http://ipindiaonline.gov.in/patentsearch/GrantedSearch/viewdoc.aspx?id=0mHccb9sR6rv6FohrBuxqA==&loc=wDBSZCsAt7zoiVrqcFJsRw==


« Previous Patent
Next Patent »
Patent Number 269226
Indian Patent Application Number 1657/KOLNP/2009
PG Journal Number 42/2015
Publication Date 16-Oct-2015
Grant Date 12-Oct-2015
Date of Filing 04-May-2009
Name of Patentee FRAUNHOFER-GESELLSCHAFT ZUR FOERDERUNG DER ANGEWANDTEN FORSCHUNG E.V.
Applicant Address HANSASTRASSE 27C, 80686 MUENCHEN
Inventors:
# Inventor's Name Inventor's Address
1 BERND EDLER HEMELINGSTRASSE 10 30419 HANNOVER
2 CHRISTIAN ERTEL GARTENSTRASSE 15 91338 LGENSDORF
3 JOHANNES HILPERT HERRNHUETTESTRASSE 46 90411 NUERNBERG
4 HARALD POPP OBERMICHELBACHERSTRASSE 18 90587 TUCHENBACH
5 RALF GEIGER MAXTORGRABEN 29 90409 NUERNBERG
PCT International Classification Number G10L 19/14
PCT International Application Number PCT/EP2007/008477
PCT International Filing date 2007-09-28
PCT Conventions:
# PCT Application Number Date of Convention Priority Country
1 10 2006 051 673.7 2006-11-02 Germany