Audio Encoder and Audio Decoder
Specification
The present invention is in the field of audio coding,
especially in the field of context based entropy coding.
Traditional audio coding concepts include an entropy coding
scheme for redundancy reduction. Typically, entropy coding
is applied to quantized spectral coefficients for frequency
domain based coding schemes or quantized time domain
samples for time domain based coding schemes. These entropy
coding schemes typically make use of transmitting a
codeword in combination with an according codebook index,
which enables a decoder to look up a certain codebook page
for decoding an encoded information word corresponding to
the transmitted codeword on the said page. In some coding
concepts, transmission of the codebook index may
nevertheless not be mandatory, for example, for cases in
which the codebook index can be determined from the context
of a symbol, which is, for example, entropy coded as
described in Meine, Edler, "Improved Quantization and
Lossless Coding for Subband Audio Coding" and Meine,
"Vektorquantisierung und kontextabhangige arithmetische
Codierung fur MPEG-4 AAC", Dissertation, Gottfried Wilhelm
Leibnitz Universitat Hannover, Hanover 2007.
For frequency or spectral domain based audio coding, a
context may describe symbols or statistical properties, as
for example, quantized spectral coefficients, which are
prior in time and/or frequency. In some of the conventional
concepts, these symbols may be available at both, the
encoder and the decoder side, and based on these symbols, a
codebook or context can be determined synchronously at both
encoder and decoder sides.

Fig. 9 shall illustrate an example of a context and its
dependencies. Fig. 9 shows a time frequency plane, in which
a number of symbols are indicated. Symbol Sn,m denotes a
symbol at time n and frequency m. Fig. 9 illustrates that
for encoding a certain symbol, its context is used to
determine the associated codebook. For example, for a
symbol Sn0,m0 this would be all symbols with
n < nO and any m, or with n = n0 and m < m0.
In practical implementations, a context is not infinite,
but limited. In the example depicted in Fig. 9, the context
for symbol S0,3 could, for example, be

For frequency based audio coding, time variant, signal
adaptive filter banks or so-called block transformations
may be used, as for example, described in Edler, B.,
"Codierung von Audiosignalen mit iiberlappender
Transformation und adaptiven Fensterfunktionen", Frequenz,
Ausgabe 43, September 1989.
In other words, frequency/time resolution changes may occur
over time within these audio coding concepts. A popular
audio coding concept is the so-called AAC (AAC = Advanced
Audio Coding), in case of which two block lengths are used,
for which are coded for example either 128 or 1024
transformed coefficients representing the frequency
components of 256 or 2048 windowed time domain samples
respectively.
These concepts allow the switching between the different
resolutions, depending on certain signal characteristics,
as for example occurrence of transients or tonality or
whether the signal is music-like or speech-like, etc. In

case of switching between different time/frequency
resolutions, as for example, between different AAC block
types, the context is not consistent. Conventional concepts
or state-of-the-art implementations may utilize resetting
of the context, i.e., it is basically switched to a state,
in which no context is available, in which a context is
built up from scratch. This approach may work out
sufficiently well, for example in AAC, since it guarantees
at least two long blocks or eight short blocks in a row,
where it can be assumed that switching occurs only seldom.
However, conventional concepts resetting the context are in
general sub-optimal in terms of coding efficiency, since
each time the context is reset, the subsequent codebook
selection is based on values, which are designed as fall-
back solutions for unknown context. Generally, sub-optimal
codebooks are then selected. The drawback in coding
efficiency may be negligible for cases in which the
switching occurs only seldom. For a scenario, however, with
more frequent switching, this leads to a significant loss
in coding efficiency. On the one hand, a more frequent
switching is strongly desired for lower data rates/sampling
rates, since especially here, an optimum adaptation of the
transform length to the signal characteristic is desired.
On the other hand, coding efficiency decreases
significantly when switching frequently.
It is the object of the present invention to provide a
concept for switching between different transform lengths
in audio coding providing an improved coding efficiency.
The object is achieved by an audio encoder according to
claim 1, a method for audio encoding according to claim 8,

an audio decoder according to claim 9, and a method for
audio decoding according to claim 16.
The present invention is based on the finding that in
context based coding such as, for example, context based
entropy coding, which can be applied to different
time/frequency resolutions, a context mapping mechanism can
be used, in case of time/frequency resolution changing over
time, thereby achieving an improved coding efficiency. It
is one finding of the present invention that when switching
between different time or frequency resolutions, contexts
for coefficients having the new resolutions can be derived
from coefficients having the old resolutions. It is one
finding of the present invention that, for example,
interpolation, extrapolation, sub-sampling, down-sampling,
up-sampling etc. can be used for context adaptation and/or
derivation when switching time/frequency resolutions in
audio coding.
Embodiments of the present invention provide a mapping
method, which maps the frequency or spectral coefficients
of a stored context, which refers to an old resolution, to
the frequency resolution of a current context or a current
frame. In other words, previous context information can be
used for codebook determination, i.e. to derive new context
information. Embodiments may therewith enable a more
frequent switching of block length and therefore a better
adaptation to signal characteristics without losing coding
efficiency.
Embodiments of the present invention will be detailed using
the accompanying figures, in which

Fig. 1 shows an embodiment of an audio encoder;
Fig. 2 shows an embodiment of an audio decoder;
Fig. 3 shows an embodiment for a context up-sampling;
Fig. 4 shows an embodiment for a context down-sampling;
Fig. 5 illustrates audio switching time and frequency
resolutions;
Fig. 6 illustrates an implementation of an embodiment;
Fig. 7a shows a flowchart of an embodiment of a method
for encoding;
Fig. 7b illustrates the general context update procedure
of an embodiment;
Fig. 7c illustrates the context update procedure of an
embodiment for resolution changes;
Fig. 8 shows a flowchart of an embodiment of a method
for decoding; and
Fig. 9 shows a state-of-the-art time frequency coding
scheme.
Fig. 1 shows an embodiment of an audio encoder 100 for
encoding segments of coefficients, the segments of
coefficients representing different time or frequency
resolutions of a sampled audio signal. The audio encoder
100 comprises a processor 110 for deriving a coding context

for a currently encoded coefficient of a current segment
based on a previously encoded coefficient of a previous
segment, the previously encoded coefficient representing a
different time or frequency resolution than the currently
encoded coefficient. The embodiment of the audio encoder
further comprises an entropy encoder 120 for entropy
encoding the current coefficient based on the coding
context to obtain an encoded audio stream.
In embodiments the coefficients may correspond to audio
samples, quantized audio samples, spectral or frequency
coefficients, scaled coefficients, transformed or filtered
coefficients etc. or any combination thereof.
In embodiments the audio encoder 100 may further comprise a
means for providing the segments of coefficients from an
audio stream, the coefficients forming a spectral
representation of an audio signal at a spectral resolution
varying among the coefficients. The means for providing the
segments may be adapted for determining the segments based
on different time domain window lengths or different audio
frames, i.e., audio frames having different lengths or
different numbers of coefficients per bandwidth, i.e. a
different spectral or frequency resolution. The means for
providing may be adapted for determining segments of 1024
and 128 time, frequency or spectral coefficients.
The processor 110 may in embodiments be adapted for
deriving the coding context based on the frequency or
spectral domain representations of the current and previous
coefficients or segments. In other words, in embodiments
the successive segments may be represented in different
time and/or frequency or spectral domains. The processor

110 may be adapted for deriving the coding context per
frequency or spectral band of the current segment, e.g.
based on neighbouring spectral coefficients of previous
segments and/or the current segment. In embodiments, the
segments may be initially determined in the time domain,
for example, by windowing an input audio stream. Based on
these time domain segments or coefficients, frequency or
spectral domain segments or coefficients may be determined
by means of transformation. The segments may be represented
in the frequency or spectral domain in terms of energy, an
amplitude and phase, an amplitude and sign, etc. per
frequency or spectral band, i.e., the segments may be sub-
divided in different frequency or spectral bands. The
processor 110 may then derive coding contexts per frequency
or spectral band in some embodiments.
The processor 110 and the entropy encoder 120 can be
configured to operate based on a down-sampling of frequency
or spectral coefficients of a previous segment when a
previous segment belonging to the context comprises a finer
spectral or frequency resolution than the current segment.
In embodiments the processor 110 and the entropy encoder
120 can be configured to operate based on an up-sampling of
frequency or spectral coefficients of a previous segment,
when a previous segment of the context comprises a coarser
spectral or frequency resolution than the current segment.
Embodiments may provide a method for encoding segments of
coefficients representing different time or frequency
resolutions of a sampled audio signal. The method may
comprise a step of deriving a coding context for a
currently encoded or current coefficient of a current
segment based on a previously encoded or previous

coefficient from a previous segment and optionally also
based on the currently encoded or current coefficient,
wherein the previously encoded or previous coefficient
represents a different time or frequency resolution than
the currently encoded or current coefficient. The method
may further comprise a step of entropy encoding the current
coefficient based on the coding context to obtain an
encoded audio stream.
Correspondingly, embodiments may comprise an audio decoder
200 of which an embodiment is depicted in Fig. 2. The audio
decoder 200 is adapted for decoding an encoded audio stream
to obtain segments of coefficients representing different
time or frequency resolutions of a sampled audio signal,
the audio decoder 200 comprises a processor 210 for
deriving a coding context for a currently decoded or
current coefficient based on a previously decoded or
previous coefficient, the previously decoded or previous
coefficient representing a different time or frequency
resolution than the currently decoded coefficient.
Furthermore, the audio decoder 200 comprises an entropy
decoder 220 for entropy decoding the current coefficient
based on the coding context and the encoded audio stream.
In embodiments, the audio decoder 200 may comprise an
entropy decoder 220, which is adapted for determining the
segments of decoded coefficients based on different time
domain window lengths or different audio frame lengths. The
entropy decoder 220 may be adapted for determining segments
of, for example, 1024 and 128, time domain samples or
frequency or spectral coefficients. Correspondingly, the
processor 210 may be adapted for deriving the coding
context based on a frequency or spectral domain

representation of coefficients of previous segments and/or
the current segment.
In embodiments, the processor 210 may be adapted for
deriving the coding context per frequency or spectral band
of the current segment, for example based on neighbouring
spectral coefficients of the previous segment or segments
and optionally from the current segment. In other words,
the segments may be processed in the frequency or spectral
domain, which can be carried out per frequency or spectral
band. Correspondingly, the processor 210 can then be
adapted for deriving a frequency or spectral band specific
context.
The entropy decoder 200 may be adapted for entropy decoding
the current coefficient based on an entropy or variable
length encoding rule.
The processor 210 may be adapted for deriving the coding
context based on a down-sampling of frequency or spectral
coefficients of a previous segment when the previous
segment comprises more coefficients per bandwidth (i.e. a
finer spectral or frequency resolution) than the current
segment. In further embodiments, the processor 210 and the
entropy encoder 220 may be configured to operate based on a
up-sampling of spectral coefficients of a previous segment,
when the previous segment comprises less coefficients per
bandwidth (i.e. a coarser spectral or frequency resolution)
than the current segment.
Consequently, embodiments may provide a method for decoding
an encoded audio stream to obtain segments of coefficients
representing decoded audio samples. The method for decoding

may comprise a step of deriving a coding context for a
currently decoded or current coefficient of a current
segment based on a previously decoded or previous
coefficient of a previous segment, the previously decoded
or previous coefficient representing a different time or
frequency resolution than the currently decoded
coefficient. Furthermore, the method may comprise a step of
entropy decoding the current coefficient based on the
coding context and the encoded audio stream. Optionally,
the method may comprise a step of determining the segments
of encoded audio coefficients from the encoded audio
stream, the segments representing different numbers of
audio coefficients.
Fig. 3 illustrates how a processor 110;210 may derive a
coding context for a current segment of Mc,new coefficients
based on a previous segment of Mc,old coefficients, wherein
the previous segment comprises a different number of audio
coefficients than the current segment. In the embodiment
depicted in Fig. 3, the number of coefficients of the
segment M determines the frequency or spectral resolution
of the segment. The embodiment may comprise a mapping
method, which maps the MC,old coefficients of a previous
segment to Mc,new coefficients having the same frequency or
spectral resolution of the context as the current segment.
Fig. 3 shows two set of coefficients within two segments,
i.e. the original previous segment 310 representing Mc,old
coefficients, Sn,0, Sn,1, Sn,2, etc., and correspondingly, the
mapped previous segment 320, which has a higher resolution,
i.e., Mc,new is greater than MC,old, representing Mc,new
coefficients, Sn,0, Sn,1, Sn,2, Sn,3, etc.

Generally, two embodiments can be distinguished, depending
on whether the resolution of the context of the current
segment is higher or lower than the resolution of the
context of the previous segment. Fig. 3 illustrates an
embodiment, in which the resolution of the previous segment
of MC,old coefficients is lower than the resolution of the
current segment of MC,new coefficients. Fig. 3 shows the
coefficients of the previous segment 310 and the symbols
for the mapped previous segment 320. From Fig. 3 can be
seen that the resolution of the current segment of Mc,new
coefficients is higher than the resolution of the previous
segment 310 having only MC,old coefficients. In one
embodiment the previous segment 310 is up-sampled to a
segment 320 of Mc,new coefficients for matching the
frequency or spectral resolution of the current segment.
This may include pure up-sampling with symbol duplication
and decimation mechanisms as for example, repeating each
value Mc,new times before decimating the resulting up-
sampled segment by keeping only 1 coefficient every MC,old,
Other interpolation or extrapolation mechanisms can also be
used.
In embodiments, mapping can be carried out for all previous
segments 310, which are needed to determine the contexts
for the current segment, e.g. at time n, in other words,
multiple previous segments may be taken into account, i.e.,
previous segments at times n-1, n-2, etc. In general,
embodiments may take multiple time slots or previous
segments into account, the number of time slots needed to
define a complete context may be different for different
implementations or embodiments.

Fig. 4 illustrates another embodiment, in which
coefficients of a previous segment 410 are down-sampled to
a segment 420 used for computing the contexts of the
current segment, i.e., in which the number of coefficients
Mc,old of the previous segment 410 is higher than the number
of coefficients Mc,new of the current segment. Fig. 4 uses a
similar illustration as Fig. 3, accordingly multiple
coefficients are shown in each segment 410 and 420. As
illustrated in Fig. 4, MC,old is greater than Mc,new.
Therefore, the Mc,old coefficients are sub-sampled, to match
the frequency or spectral resolution of the current segment
of Mc,new coefficients, i.e. in embodiments previous
segments having a higher resolution may be sub-sampled to
match the resolution of the current segment having a lower
resolution. In embodiments this may include pure down-
sampling with coefficient duplication and decimation
mechanisms as for example, repeating each value Mc,new times
before decimating the resulting up-sampled segment by
keeping only 1 coefficient every Mc,old. In other
embodiments, filter operations, as for example averaging of
two or multiple adjacent values may be taken into account.
Fig. 5 illustrates another embodiment, in which switching
between different resolutions is carried out. Fig. 5 shows
a time/frequency plane, in which three subsequent segments
of audio coefficients are shown, namely 510, 520 and 530.
Each of the segments 510, 520 and 530 corresponds to a
single set of coefficients. In the embodiment illustrated
in Fig. 5, it is assumed that the second segment 520 is
twice as long a the first and third segments 510 and 530.
This may be achieved by using different windows when
segmenting in the time domain, as for example done in AAC.
In the embodiment illustrated in Fig. 5, it is assumed that

the sampling rate remains constant, in other words, the
longer second segment 520 comprises twice as many audio
coefficients per bandwidth than the first or third segment
510 or 530.
Fig. 5 shows that in this case the resolution in the
frequency or spectral domain scales with the extent of the
segment in the time domain. In other words, the shorter the
window in the time domain, the lower the resolution in the
frequency or spectral domain. When evaluating contexts for
encoding the samples in the frequency or spectral domain,
Fig. 5 shows that the coding needs to have a higher
resolution version of the segment 510 when encoding the
second segment 520 as in the example considered, a double
resolution of the segment 510 has to be derived. In other
embodiments, especially when using other time-frequency
transformations or filterbanks, other relations between the
time domain and frequency domain resolutions may result.
According to an embodiment, the coefficients encoded during
the first segment 510 provides a basis for determining the
context for the second segment 520, for example, by means
of an intermediate up-sampling. In other words, the context
content coming from the first segment 510 may be obtained
by an up-sampling of the first segment 510, for example, in
terms of interpolation or extrapolation, in order to derive
the context for the second segment 520, having a higher
resolution.
As shown in Fig. 5, when switching from the second segment
520 to the third segment 530, the context constituent
element has to change as well, as the resolution has now
decreased. According to an embodiment, the coefficients

encoded during the second segment 520 may be used in order
to derive the context for the third segment, by means of an
intermediate down-sampling. This may be carried out, for
example in terms of averaging or by simply using only every
second value or other measures for down-sampling.
Embodiments achieve the advantage of an increased coding
efficiency, by taking into account the past context derived
from previous segments even when changes in the resolution
or window length occur. Context constituent elements can be
adapted to new resolutions, in terms of up or down-
sampling, for example, with inter- or extrapolation,
filtering or averaging, etc.
In the following a more specific embodiment will be
presented in terms of spectral noiseless coding. Spectral
noiseless coding may be used for further reducing the
redundancy of a quantized spectrum in audio coding.
Spectral noiseless coding can be based on an arithmetic
coding in conjunction with dynamically context adaptation.
Noiseless coding can be based on quantized spectral values
and may use context dependent cumulative frequency tables
derived from, for example, four previously decoded
neighbouring tuples. Fig. 6 illustrates another embodiment.
Fig. 6 shows a time frequency plane, wherein along the time
axis three time slots are indexed n, n-1 and n-2.
Furthermore, Fig. 6 illustrates four frequency or spectral
bands which are labelled by m-2, m-1, m and m+1. Fig. 6
shows within each time-frequency slot boxes, which
represent tuples of samples to be encoded or decoded. Three
different types of tuples are illustrated in Fig. 6, in
which boxes having a dashed or dotted border indicate

remaining tuples to be encoded or decoded, white boxes
having a solid border indicate previously encoded or
decoded tuples and grey boxes with a solid border indicate
previously en/decoded tuples, which are used to determine
the context for the current tuple to be encoded or decoded.
Note that the previous and current segments referred to in
the above described embodiments may correspond to a tuple
in the present embodiment, in other words, the segments may
be processed bandwise in the frequency or spectral domain.
As illustrated in Fig. 6, tuples or segments in the
neighbourhood of a current tuple (i.e. in the time and the
frequency or spectral domain) may be taken into account for
deriving a context. Cumulative frequency tables may then be
used by the arithmetic coder to generate a variable length
binary code. The arithmetic coder may produce a binary code
for a given set of symbols and their respective
probabilities. The binary code may be generated by mapping
a probability interval, where the set of symbols lies, to a
codeword. The arithmetic coder may correspond to the
entropy encoder 120, respectively the entropy decoder 220
in the above described embodiments.
In the present embodiment context based arithmetic coding
may be carried out on the basis of 4-tuples (i.e. on four
spectral coefficient indices), which are also labelled
q(n,m), representing the spectral coefficients after
quantization, which are neighboured in the frequency or
spectral domain and which are entropy coded in one step.
According to the above description, coding may be carried
out based on the coding context. As indicated in Fig. 6,
additionally to the 4-tuple, which is coded (i.e. the
current segment) four previously coded 4-tuples are taken

into account in order to derive the context. These four 4-
tuples determine the context and are previous in the
frequency and/or previous in the time domain.
Fig. 7a shows a flow-chart of a USAC (USAC = Universal
Speech and Audio Coder) context dependent arithmetic coder
for the encoding scheme of spectral coefficients. The
encoding process depends on the current 4-tuple plus the
context, where the context is used for selecting the
probability distribution of the arithmetic coder and for
predicting the amplitude of the spectral coefficients. In
Fig. 7a the box 705 represents context determination, which
is based on tO, tl, t2 and t3 corresponding to q(n-l, m) ,
q(n,m-l), q (n-l,m-l) and q (n-l,m+l) , that is the grey
boxes with solid borders in Fig. 6.
Generally, in embodiments the entropy encoder can be
adapted for encoding the current segment in units of a 4-
tuple of spectral coefficients and for predicting an
amplitude range of the 4-tuple based on the coding context.
In the present embodiment the encoding scheme comprises
several stages. First, the literal codeword is encoded
using an arithmetic coder and a specific probability
distribution. The codeword represents four neighbouring
spectral coefficients (a,b,c,d), however, each of a, b, c,
d is limited in range:

Generally, in embodiments the entropy encoder 120 can be
adapted for dividing the 4-tuple by a predetermined factor
as often as necessary to fit a result of the division in
the predicted range or in a predetermined range and for

encoding a number of divisions necessary, a division
remainder and the result of the division when the 4-tuple
does not lie in the predicted range, and for encoding a
division remainder and the result of the division
otherwise.
In the following, if the term (a,b,c,d), i.e. any
coefficient a, b, c, d, exceeds the given range in this
embodiment, this can in general be considered by dividing
(a,b,c,d) as often by a factor (e.g. 2 or 4) as necessary,
for fitting the resulting codeword in the given range. The
division by a factor of 2 corresponds to a binary shifting
to the right-hand side, i.e. (a,b,c,d)» 1. This diminution
is done in an integer representation, i.e. information may
be lost. The least significant bits, which may get lost by
the shifting to the right, are stored and later on coded
using the arithmetic coder and a uniform probability
distribution. The process of shifting to the right is
carried out for all four spectral coefficients (a,b,c,d).
In general embodiments, the entropy encoder 120 can be
adapted for encoding the result of the division or the 4-
tuple using a group index ng, the group index ng referring
to a group of one or more codewords for which a probability
distribution is based on the coding context, and an element
index ne in case the group comprises more than one
codeword, the element index ne referring to a codeword
within the group and the element index can be assumed
uniformly distributed, and for encoding the number of
divisions by a number of escape symbols, an escape symbol
being a specific group index ng only used for indicating a
division and for encoding the remainders of the divisions
based on a uniform distribution using an arithmetic coding

rule. The entropy encoder 120 can be adapted for encoding a
sequence of symbols into the encoded audio stream using a
symbol. alphabet comprising the escape symbol, and group
symbols corresponding to a set of available group indices,
a symbol alphabet comprising the corresponding element
indices, and a symbol alphabet comprising the different
values of the remainders.
In the embodiment of Fig. 7a, the probability distribution
for encoding the literal codeword and also an estimation of
the number of range-reduction steps can be derived from the
context. For example, all codewords, in a total 84 = 4096,
span in total 544 groups, which consist of one or more
elements. The codeword can be represented in the bitstream
as the group index ng and the group element ne. Both values
can be coded using the arithmetic coder, using certain
probability distributions. In one embodiment the
probability distribution for ng may be derived from the
context, whereas the probability distribution for ne may be
assumed to be uniform. A combination of ng and ne may
unambiguously identify a codeword. The remainder of the
division, i.e. the bit-planes shifted out, may be assumed
to be uniformly distributed as well.
In Fig. 7a, in step 710, the 4-tuple q(n,m), that is
(a,b,c,d) or the current segment is provided and a
parameter lev is initiated by setting it to 0.
In step 715 from the context, the range of (a,b,c,d) is
estimated. According to this estimation, (a,b,c,d) may be
reduced by lev0 levels, i.e. divided by a factor of 2lev0.
The levO least significant bitplanes are stored for later
usage in step 750.

In step 720 it is checked whether (a,b,c,d) exceeds the
given range and if so, the range of (a,b,c,d) is reduced by
a factor of 4 in step 725. In other words, in step 725
(a,b,c,d) are shifted by 2 to the right and the removed
bitplanes are stored for later usage in step 750.
In order to indicate this reduction step, ng is set to 544
in step 730, i.e. ng = 544 serves as an escape codeword.
This codeword is then written to the bitstream in step 755,
where for deriving the codeword in step 730 an arithmetic
coder with a probability distribution derived from the
context is used. In case this reduction step was applied
the first time, i.e. if lev==lev0, the context is slightly
adapted. In case the reduction step is applied more than
once, the context is discarded and a default distribution
is used further on. The process then continues with step
720.
If in step 720 a match for the range is detected, more
specifically if (a,b,c,d) matches the range condition,
(a,b,c,d) is mapped to a group ng, and, if applicable, the
group element index ne. This mapping is unambiguously, that
is (a,b,c,d) can be derived from ng and ne. The group index
ng is then coded by the arithmetic coder, using a
probability distribution arrived for the adapted/discarded
context in step 735. The group index ng is then inserted
into the bitstream in step 755. In a following step 740,
it is checked whether the number of elements in the group
is larger than 1. If necessary, that is if the group
indexed by ng consists of more than one element, the group
element index ne is coded by the arithmetic coder in step

745, assuming a uniform probability distribution in the
present embodiment.
Following step 745, the element group index ne is inserted
into the bitstream in step 755. Finally, in step 750, all
stored bitplanes are coded using the arithmetic coder,
assuming a uniform probability distribution. The coded
stored bitplanes are then also inserted into the bitstream
in step 755.
In embodiments the entropy decoder 220 can be adapted for
decoding a group index ng from the encoded audio stream
based on a probability distribution derived from the coding
context, wherein the group index ng represents a group of
one or more codewords, and for, based on a uniform
probability distribution, decoding an element index ne from
the encoded audio stream if the group index ng indicates a
group comprising more than one codeword, and for deriving a
4-tuple of spectral coefficients of the current segment
based on the group index ng and the element index ne,
thereby obtaining the spectral domain representation in
tuples of spectral coefficients.
In embodiments the entropy decoder 220 can be adapted for
decoding a sequence of symbols from the encoded audio
stream based on the probability distribution derived from
the coding context using a symbol alphabet comprising an
escape symbol and group symbols corresponding to a set of
available group indices ng, for deriving a preliminary 4-
tuple of spectral coefficients based on an available group
index ng to which a group symbol of the sequence of symbols
corresponds, and based on the element index ne, and for
multiplying the preliminary 4-tuple with a factor depending

on a number of escape symbols in the sequence of symbols to
obtain the tuple of spectral coefficients.
The entropy decoder 220 may be further adapted for decoding
a remainder from the encoded audio stream based on a
uniform probability distribution using an arithmetic coding
rule and for adding the remainder to the multiplied
preliminary 4-tuple to obtain the 4-tuple of spectral
coefficients.
The entropy decoder 220 can be adapted for multiplying the
4-tuple with a predetermined factor as often as an escape
symbol is decoded from the encoded audio stream, an escape
symbol being a specific group index ng only used for
indicating a multiplication, and for decoding a remainder
from the encoded audio stream based on a uniform
probability distribution using an arithmetic coding rule,
the entropy decoder 220 can be further adapted for adding
the remainder to the multiplied 4-tuple to obtain the
current segment.
In the following an embodiment of a USAC context dependent
arithmetic coder decoding scheme will be described.
Corresponding to the above embodiment of the encoding
scheme, 4-tuples corresponding to quantized spectral
coefficients, which are noiselessly coded, are considered.
Furthermore, it is assumed that the 4-tuples are
transmitted starting from the lowest frequency or spectral
coefficient and progressing to the highest frequency or
spectral coefficient. The coefficients may, for example,
correspond to AAC coefficients, which are stored in an
array, and the order of transmission of the noiseless
coding codewords is supposed to be such that when they are

decoded in the order received and stored in the array, bin
is the most rapidly incrementing index and g is the most
slowly incrementing index. Within a codeword, the order of
decoding is a,b,c,d.
Fig. 7b illustrates the general context update procedure of
an embodiment. Details on the context adaptation according
to a bit depth prediction mechanism will be considered in
the present embodiment. Fig. 7b shows a plane 760
displaying the possible range of a 4-tuple (a,b,c,d) in
terms of bitplanes. The bit depth, i.e. number of bit
planes needed to represent a 4-tuple, can be predicted by
the context of the current 4-tuple by means of the
calculation of the variable called levO, which is also
indicated in Fig. 7b. The 4-tuple is then divided by 2lev0,
i.e. Iev=lev0 bit planes are removed and stored for later
usage according to the above-described step 715.
If the 4-tuple is in the range -5lev0+2 another escape symbol is coded according to
step 770 in Fig. 7b the context is completely reset, cf.
772, discarded as in step 730 in Fig. 7a, respectively. No
context adaptation is used further because it is considered
not relevant for the current 4-tuple coding. The default
probability model, the one used when no context is
available, is then used for the future ng symbols, which is
indicated by steps 774 and 776 in Fig. 7b. The process is
then repeated for other tuples.
To summarize, the context adaptation is a mechanism in
order to reduce the context significance in the context-
adaptive coding. The context adaptation can be triggered
when the predicted levO and the actual lev mismatch. It is
easily detected by the number of coded escape symbols
(ng=544), cf. 766 and 770 in Fig. 7b, and can therefore be
carried out at the decoder in a similar way as well.
The context adaptation can be done by triggering a flag in
the context state representation t. The value t is
calculated by the function get_state(), as levO, by using
the context derived from the previous and/or current frame
or segment of the current 4-tuple, which is stored in a
table q[][]. The state of the context can e.g. be
represented by 24 bits. There are 1905800 states possible

in an embodiment. These states can be represented by only
21 bits. The 23th and 24th bits of t are reserved for
adapting the context state. According to the values of the
23th and the 24th bits, the get_pk() will output different
probability distribution models. In one embodiment, the
23th bit of t may be set to one when the 4-tuple is divided
by 4 after being previously divided by lev0, i.e.
lev==lev0+2.
In consequence the mapping between the context state t and
the probability distribution model pki, is different for
lev==lev0+2 than for lev==lev0. The mapping between the
context state t and the model pki is pre-defined during a
training phase by performing optimizations on the overall
statistics of the training sequence. When lev>lev0+2, the
context and t may be set to zero. Get_pk() outputs then the
default probability distribution model pki, which
corresponds to t=0.
In the following the details of a context mapping in one
embodiment will be described. The context mapping is the
first operation done in context-adaptive coding after the
eventual reset of the context in the present embodiment. It
is done in two steps.
First, before the coding, the context table qs[] of size
previous_lg/4, saved at the previous frame, is mapped in a
context table q[0][] of size lg/4 corresponding to the size
of the current frame. The mapping is done in the
arith_map_context function, which is exemplified by the
following psuedo-code:



As can be seen from the pseudo-code, the mapping may not
exactly the same for all coding strategies. In the present
embodiment, mapping differs when AAC (Advanced Audio
Coding) is used (core_mode==0) for a coefficient from when
TCX (Transform based Coding) is used (core_mode==1). One of
the differences comes from how the bounds of the tables are
handled. In AAC, the mapping may start from index 0 (first
value in the table) , while for TCX it may start from the
index 1 (second value in table) knowing that the first
value is always set to 'unknown' (specific state used to
reset the context) . The ratio of previous_lg over lg will
determine the order of the upsampling (ratio<1) or the
downsampling (ratio>1), which will be performed in present

embodiment. Fig. 7c illustrates the case of TCX when
converting from a saved context table of size 1024/4, cf.
left hand side 780 of Fig. 7c, to a size of 512/4, cf.
right hand side 782 of Fig. 7c. It can be seen, that while
for the current context table 782 stepwise increments of 1
are used, for the saved context table 780 stepwise
increments of the above described ratio are used.
Fig. 7c illustrates the context update procedure of an
embodiment for resolution changes. Once the mapping is
done, the context-adaptive coding is performed. At the end
of the coding, the current frame elements are saved in
table qs[] for the next frame. It may be done in
arith_update_context() , which is exemplified by the
following pseudo-code:



In the present embodiment, the saving is done differently
according to the core coder (AAC or TCX) . In TCX the
context is always saved in a table qs[] of 1024/4 values.
This additional mapping can be done because of the closed-
loop decision of the AMR-WB+ (Adaptive Multirate WideBand
Codec). In the close-loop decision several copy procedures
of the coder states are needed for testing each possible
combination of TCXs and ACELP (Arithmetic Coded Excited
Linear Prediction). The state copy is easier to implement
when all the TCX modes share the same size for the table
qs[]. A mapping is then used to convert systematically
from lg/4 to 1024/4. On the other hand, AAC saves only the
context and performs no mapping during this phase.
Fig. 8 illustrates a flowchart of the embodiment of the
decoding scheme. In step 805, corresponding to step 705,
the context is derived on the basis of t0, t1, t2 and t3.
In step 810, the first reduction level lev0 is estimated

from the context, and the variable lev is set to lev0. In
the following step 815, the group ng is read from the
bitstream and the probability distribution for decoding ng
is derived from the context. In step 815, the group ng can
then be decoded from the bitstream.
In step 820 it is determined whether the ng equals 544,
which corresponds to the escape value. If so, the variable
lev can be increased by 2 before returning to step 815. In
case this branch is used for the first time, i.e., if
lev==lev0, the probability distribution respectively the
context can be accordingly adapted, respectively discarded
if the branch is not used for the first time, in line with
the above described context adaptation mechanism, cf. Figs.
7b and 7c. In case the group index ng is not equal to 544
in step 820, in a following step 825 it is determined
whether the number of elements in a group is greater than
1, and if so, in step 830, the group element ne is read and
decoded from the bitstream assuming a uniform probability
distribution. The element index ne is derived from the
bitstream using arithmetic coding and a uniform probability
distribution.
In step 835 the literal codeword (a,b,c,d) is derived from
ng and ne, for example, by a look-up process in the tables,
for example, refer to dgroups[ng] and acod_ne[ne].
In step 840 for all lev missing bitplanes, the planes are
read from the bitstream using arithmetic coding and
assuming a uniform probability distribution. The bitplanes
can then be appended to (a,b,c,d) by shifting (a,b,c,d) to
the left and adding the bitplane bp: ( (a,b, c, d) «=1) | =bp.
This process may be repeated lev times.

Finally in step 845 the 4-tuple q(n,m), i.e. (a,b, c, d) can
be provided.
In the following pseudo codes and implementation details
are provide according to an embodiment. The following
definitions will be used.
(a,b,c,d) 4-tuple to decode
ng Group index of the most significant 2-bits-wise plane
of the 4-tuple, where 0 <= ng <= 544. The last value
544 corresponds to the escape symbol, ARITH_ESCAPE.
ne Element index within a group, ne lies between 0 and
the cardinal of each group mm. The maximum number of
elements within a group is 73.
lev Level of the remaining bit-planes. It corresponds to
number the bit planes less significant than the most
significant 2 bits-wise plane.
egroups [a][b][c][d] Group index table. It permits to map
the most significant 2bits-wise plane of the 4-
tuple (a,b,c,d) into the 544 groups.
mm Cardinal of the group
og Offset of the group
dgroups[] Maps the group index ng to the cardinal of each
group mm (first 8 bits) and the offset of the
group og in dgvectors[] (last 8 bits).
dgvectors[] Map the offset of the group og and the index
of the element ne to the most significant 2
bits-wise plane of the 4-tuple (a,b,c,d).

arith_cf_ng_hash[] Hash table mapping the state of context
to a cumulative frequencies table index
pki.
arith_cf_ng[pki] [545] Models of the cumulative
frequencies for the group index
symbol ng.
arith_cf_ne [ ] Cumulative frequencies for element
index symbol ne.
r Bit plane of the 4-tuple less significant than
the most significant 2-bits wise plane.
arith_cf_r [] Cumulative frequencies for the least
significant bit-planes symbol r
In the following the decoding process is considered first.
Four-tuples quantized spectral coefficients are noiselessly
coded and transmitted starting from the lowest-frequency or
spectral coefficient and progressing to the highest-
frequency or spectral coefficient. The coefficients from
AAC are stored in the array x_ac_quant[g][win][sfb][bin],
and the order of transmission of the noiseless coding
codewords is such that when they are decoded in the order
received and stored in the array, bin is the most rapidly
incrementing index and g is the most slowly incrementing
index. Within a codeword the order of decoding is a, b, c,
d. The coefficient from the TCX are stored directly in the
array x_tcx_invquant [win] [bin] , and the order of the
transmission of the noiseless coding codewords is such that
when they are decoded in the order received and stored in
the array, bin is the most rapidly incrementing index and
win is the most slowly incrementing index. Within a
codeword the order of decoding is a, b, c, d. First, the
flag arith_reset_flag determines if the context must be
reset. If the flag is TRUE the following function is
called:

The noiseless decoder outputs 4-tuples of signed quantized
spectral coefficients. At first the state of the context is
calculated based on the four previously decoded groups
surrounding the 4-tuple to decode. The state of the context
is given by the function arith_get_context():



Once the state is known the group to which belongs the most
significant 2-bits wise plane of 4-tuple is decoded using
the arith_decode() fed with the appropriated cumulative


Then the arith_decode() function is called with the
cumulative frequencies table corresponding to the index
return by the arith_get_pk(). The arithmetic coder is an

While the decoded group index ng is the escape symbol,
ARITH_ESCAPE, an additional group index ng is decoded and
the variable lev is incremented by two. Once the decoded
group index is not the escape symbol, ARITH_ESCAPE, the
number of elements, mm, within the group and the group
offset, og, are deduced by looking up to the table
dgroups[]:


The element index ne is then decoded by calling
arith_decode() with the cumulative frequencies table
(arith_cf_ne+( (mm* (mm-1) ) »1) [] . Once the element index is
decoded the most significant 2-bits wise plane of the 4-
tuple can be derived with the table dgvector[]:

The remaining bit planes are then decoded from the most
significant to the lowest significant level by calling lev
times arith_decode() with the cumulative frequencies table
arith_cf_r []. The decoded bit plane r permits to refine
the decode 4-tuple by the following way:

Once the 4-tuple (a,b,c,d) is completely decoded the
context tables q and qs are updated by calling the function
arith_update_context().
arith_update_context(a,b,c,d, i, lg)



Depending on certain implementation requirements of the
inventive methods, the inventive methods may be implemented
in hardware or in software. The implementation can be
formed using a digital storage medium, in particular, a
disk, a DVD, or a CD, having an electronically readable
control signal stored thereon, which cooperates with the
programmable computer, such that the inventive methods are
performed. Generally, the present invention is therefore a
computer program product with a program code for a machine-
readable carrier, the program code being operative for
performing the inventive methods when the computer program
runs on a computer. In other words, the inventive methods
are, therefore, a computer program having a program code
for performing at least one of the inventive methods when
the computer program runs on a computer.

We Claim:
1. An audio encoder (100) for encoding a sequence of
segments of coefficients, the segments being
subsequent to each other in time, the audio encoder
(100) comprising
a means for providing the sequence of segments of
coefficients from an audio stream representing a
sampled audio signal by using different transform
lengths such that segments of coefficients for which
different transform lengths are used, spectrally
represent the sampled audio signal at different
frequency resolutions;
a processor (110) for deriving a coding context for a
currently encoded coefficient of a current segment
based on a previously encoded coefficient of a
previous segment, the previous and current segments
corresponding to different frequency resolutions and
different transform lengths, respectively; and
an entropy encoder (120) for entropy encoding the
current coefficient based on the coding context to
obtain an encoded audio stream.
2. The audio encoder (100) of claim 1, wherein the
entropy encoder (120) is adapted for encoding the
current coefficient in units of a tuple of spectral
coefficients and for predicting a range of the tuple
based on the coding context.

3. The audio encoder (100) of claim 2, wherein the
entropy encoder (120) is adapted for dividing the
tuple by a predetermined factor as often as necessary
to fit a result of the division in a predetermined
range and for encoding a number of divisions
necessary, a division remainder and the result of the
division when the tuple does not lie in the predicted
range, and for encoding a division remainder and the
result of the division otherwise.
4. The audio encoder (100) of claim 3, wherein the
entropy encoder (120) is adapted for encoding the
result of the division or the tuple using a group
index, the group index referring to a group of one or
more codewords for which a probability distribution is
based on the coding context, and, based on a uniform
probability distribution, an element index in case the
group comprises more than one codeword, the element
index referring to a codeword within the group, and
for encoding the number of divisions by a number of
escape symbols, an escape symbol being a specific
group index only used for indicating a division, and
for encoding the remainders of the divisions based on
a uniform probability distribution using an arithmetic
coding rule.
5. The audio encoder (100) of claim 4, wherein the
entropy encoder (120) is adapted for encoding a
sequence of symbols into the encoded audio stream
using a symbol alphabet comprising the escape symbol,
and group symbols corresponding to a set of available
group indices, a symbol alphabet comprising the

corresponding element indices, and a symbol alphabet
comprising the different values of the remainders.
6. The audio encoder (100) of one of the claims 1 to 5,
wherein the processor (110) and the entropy encoder
(120) are configured to operate based on a down-
sampling of spectral coefficients of the previous
segment, when the previous segment displays a finer
spectral resolution than the current segment and/or
wherein the processor (110) and the entropy encoder
(120) are configured to operate based on an up-
sampling of spectral coefficients of the previous
segment, when the previous segment displays a coarser
spectral resolution than the current segment.
7. Method for encoding a sequence of segments of
coefficients, the segments being subsequent to each
other in time, the method comprising the steps of
providing the sequence of segments of coefficients
from an audio stream representing a sampled audio
signal by using different transform lengths such that
segments of coefficients for which different transform
lengths are used, spectrally represent the sampled
audio signal at different frequency resolutions;
deriving a coding context for a currently encoded
coefficient of a current segment based on a previously
encoded coefficient of a previous segment, the
previous and current segments corresponding to
different frequency resolutions and different
transform lengths, respectively; and

entropy encoding the current coefficient based on the
coding context to obtain an encoded audio stream.
8. An audio encoder (100) for encoding segments of
coefficients, the segments of coefficients
representing different time or frequency resolutions
of a sampled audio signal, the audio encoder (100)
comprising
a means for providing the segments of coefficients
from an audio stream, the coefficients forming a
spectral representation of an audio signal represented
by the audio stream at a spectral resolution varying
among the segments;
a processor (110) for deriving a coding context for a
currently encoded coefficient of a current segment
based on a previously encoded coefficient of a
previous segment, the previously encoded coefficient
representing a different time or frequency resolution
than the currently encoded coefficient; and
an entropy encoder (120) for entropy encoding the
current coefficient based on • the coding context to
obtain an encoded audio stream,
wherein the entropy encoder (120) is adapted for
encoding the current coefficient in units of a tuple
of spectral coefficients and for predicting a range of
the tuple based on the coding context,
wherein the entropy encoder (120) is adapted for
dividing the tuple by a predetermined factor as often

as necessary to fit a result of the division in a
predetermined range and for encoding a number of
divisions necessary, a division remainder and the
result of the division when the tuple does not lie in
the predicted range, and for encoding a division
remainder and the result of the division otherwise,
and
wherein the entropy encoder (120) is adapted for
encoding the result of the division or the tuple using
a group index, the group index referring to a group of
one or more codewords for which a probability
distribution is based on the coding context, and,
based on a uniform probability distribution, an
element index in case the group comprises more than
one codeword, the element index referring to a
codeword within the group, and for encoding the number
of divisions by a number of escape symbols, an escape
symbol being a specific group index only used for
indicating a division, and for encoding the remainders
of the divisions based on a uniform probability
distribution using an arithmetic coding rule.
9. An audio decoder (200) for decoding an encoded audio
stream representing a sampled audio signal to obtain a
sequence of segments of coefficients being subsequent
to each other in time and representing the sampled
audio signal by using different transform lengths such
that segments of coefficients for which different
transform lengths are used, spectrally represent the
sampled audio signal at different frequency
resolutions, comprising

a processor (210) for deriving a coding context for a
currently decoded coefficient of a current segment
based on a previously decoded coefficient of a
previous segment, the previous and current segments
corresponding to different frequency resolutions and
different transform lengths, respectively; and
an entropy decoder (220) for entropy decoding the
current coefficient based on the coding context and
the encoded audio stream.
10. The audio decoder (200) of claims 9, wherein the
processor is adapted for deriving the coding context
per spectral band for the current coefficient, based
on neighbouring spectral coefficients previously
decoded in the previous segment and optionally in the
present segment.
11. The audio decoder (200) of claim 10, wherein the
entropy decoder (220) is adapted for decoding a group
index from the encoded audio stream based on a
probability distribution derived from the coding
context, wherein the group index represents a group of
one or more codewords, and for, based on a uniform
probability distribution, decoding an element index
from the encoded audio stream if the group index
indicates a group comprising more than one codeword,
and for deriving a tuple of spectral coefficients of
the current segment based on the group index and the
element index, thereby obtaining the spectral domain
representation in tuples of spectral coefficients.

12. The audio decoder (200) of claim 11, wherein the
entropy decoder (220) is adapted for decoding a
sequence of symbols from the encoded audio stream
based on the probability distribution derived from the
coding context using a symbol alphabet comprising an
escape symbol and group symbols corresponding to a set
of available group indices, for deriving a preliminary
tuple of spectral coefficients based on an available
group index to which a group symbol of the sequence of
symbols corresponds, and based on the element index,
and for multiplying the preliminary tuple with a
factor depending on a number of escape symbols in the
sequence of symbols to obtain the tuple of spectral
coefficients.
13. The audio decoder (200) of claim 12, wherein the
entropy decoder (220) is adapted for decoding a
division remainder from the encoded audio stream based
on a uniform probability distribution using an
arithmetic coding rule and for adding the remainder to
the multiplied preliminary tuple to obtain the tuple
of spectral coefficients.
14. The audio decoder (200) of one of the claims to 13,
wherein the processor (210) and the entropy encoder
(220) are configured to operate based on a down-
sampling of spectral coefficients of the previous
segment, when the previous segment displays a finer
spectral resolution than the current segment and/or
wherein the processor (210) and the entropy encoder
(220) are configured to operate based on an up-
sampling of spectral coefficients of the previous

segment, when the previous segment displays a coarser
spectral resolution than the current segment.
15. An audio decoder (200) for decoding an encoded audio
stream to obtain segments of coefficients representing
different time or frequency resolutions of a the
sampled audio signal, comprising
a processor (210) for deriving a coding context for a
currently decoded coefficient of a current segment
based on a previously decoded coefficient of a
previous segment, the previously decoded coefficient
representing a different time or frequency resolution
than the currently decoded coefficient; and
an entropy decoder (220) for entropy decoding the
current coefficient based on the coding context and
the encoded audio stream,
wherein the processor (210) is adapted for deriving
the coding context based on the previous coefficient,
the coefficients forming a spectral representation of
an audio signal represented by the audio stream at a
spectral resolution varying among the segments,
wherein the processor is adapted for deriving the
coding context for the current coefficient, based on
neighbouring spectral coefficients previously decoded
in the previous segment and optionally in the present
segment, and
wherein the entropy decoder (220) is adapted for
decoding a group index from the encoded audio stream

based on a probability distribution derived from the
coding context, wherein the group index represents a
group of one or more codewords, and for, based on a
uniform probability distribution, decoding an element
index from the encoded audio stream if the group index
indicates a group comprising more than one codeword,
and for deriving a tuple of spectral coefficients of
the current segment based on the group index and the
element index, thereby obtaining the spectral domain
representation in tuples of spectral coefficients.
16. A method for decoding an encoded audio stream
representing a sampled audio signal to obtain a
sequence of segments of coefficients being subsequent
to each other in time and representing the sampled
audio signal by using different transform lengths such
that segments of coefficients for which different
transform lengths are used, spectrally represent the
sampled audio signal at different frequency
resolutions, comprising the steps of
deriving a coding context for a currently decoded
coefficient of a current segment based on a previously
decoded coefficient of a previous segment, the
previous and current segments corresponding to
different frequency resolutions and different
transform lengths, respectively; and
entropy decoding the current coefficient based on the
coding context and the encoded audio stream.

17. A computer program having a program code for
performing one of the methods of claims 7 or 15, when
the program code runs on a computer or a processor

An audio encoder (100) for encoding segments of
coefficients, the segments of coefficients representing
different time or frequency resolutions of a sampled audio
signal, the audio encoder (100) comprising a processor
(110) for deriving a coding context for a currently encoded
coefficient of a current segment based on a previously
encoded coefficient of a previous segment, the previously
encoded coefficient representing a different time or
frequency resolution than the currently encoded
coefficient. The audio encoder (100) further comprises an
entropy encoder (120) for entropy encoding the current
coefficient based on the coding context to obtain an
encoded audio stream.