Watermark Signal Provision and Watermark Embedding
Description
Technical Field
The present invention relates to a watermark signal provider for providing a watermark
signal and watermark embedding using the watermark signal.
Background of the Invention
In many technical applications, it is desired to include an extra information into an
information or signal representing useful data or "main data" like, for example, an audio
signal, a video signal, graphics, a measurement quantity and so on. In many cases, it is
desired to include the extra information such that the extra information is bound to the
main data (for example, audio data, video data, still image data, measurement data, text
data, and so on) in a way that it is not perceivable by a user of said data. Also, in some
cases it is desirable to include the extra data such that the extra data are not easily
removable from the main data (e.g. audio data, video data, still image data, measurement
data, and so on).
This is particularly true in applications in which it is desirable to implement a digital rights
management. However, it is sometimes simply desired to add substantially unperceivable
side information to the useful data. For example, in some cases it is desirable to add side
information to audio data, such that the side information provides an information about the
source of the audio data, the content of the audio data, rights related to the audio data and
so on.
For embedding extra data into useful data or "main data", a concept called "watermarking"
may be used. Watermarking concepts have been discussed in the literature for many
different kinds of useful data, like audio data, still image data, video data, text data, and so
on.
In the following, some references will be given in which watermarking concepts are
discussed. However, the reader's attention is also drawn to the wide field of textbook
literature and publications related to the watermarking for further details.

DE 196 40 814 C2 describes a coding method for introducing a non-audible data signal
into an audio signal and a method for decoding a data signal, which is included in an audio
signal in a non-audible form. The coding method for introducing a non-audible data signal
into an audio signal comprises converting the audio signal into the spectral domain. The
coding method also comprises determining the masking threshold of the audio signal and
the provision of a pseudo noise signal. The coding method also comprises providing the
data signal and multiplying the pseudo noise signal with the data signal, in order to obtain
a frequency-spread data signal. The coding method also comprises weighting the spread
data signal with the masking threshold and overlapping the audio signal and the weighted
data signal.
In addition, WO 93/07689 describes a method and apparatus for automatically identifying
a program broadcast by a radio station or by a television channel, or recorded on a
medium, by adding an inaudible encoded message to the sound signal of the program, the
message identifying the broadcasting channel or station, the program and/or the exact date.
In an embodiment discussed in said document, the sound signal is transmitted via an
analog-to-digital converter to a data processor enabling frequency components to be split
up, and enabling the energy in some of the frequency components to be altered in a
predetermined manner to form an encoded identification message. The output from the
data processor is connected by a digital-to-analog converter to an audio output for
broadcasting or recording the sound signal. In another embodiment discussed in said
document, an analog bandpass is employed to separate a band of frequencies from the
sound signal so that energy in the separated band may be thus altered to encode the sound
signal.
US 5, 450,490 describes apparatus and methods for including a code having at least one
code frequency component in an audio signal. The abilities of various frequency
components in the audio signal to mask the code frequency component to human hearing
are evaluated and based on these evaluations an amplitude is assigned to the code
frequency component. Methods and apparatus for detecting a code in an encoded audio
signal are also described. A code frequency component in the encoded audio signal is
detected based on an expected code amplitude or on a noise amplitude within a range of
audio frequencies including the frequency of the code component.
WO 94/11989 describes a method and apparatus for encoding/decoding broadcast or
recorded segments and monitoring audience exposure thereto. Methods and apparatus for
encoding and decoding information in broadcasts or recorded segment signals are
described. In an embodiment described in the document, an audience monitoring system

encodes identification information in the audio signal portion of a broadcast or a recorded
segment using spread spectrum encoding. The monitoring device receives an acoustically
reproduced version of the broadcast or recorded signal via a microphone, decodes the
identification information from the audio signal portion despite significant ambient noise
and stores this information, automatically providing a diary for the audience member,
which is later uploaded to a centralized facility. A separate monitoring device decodes
additional information from the broadcast signal, which is matched with the audience diary
information at the central facility. This monitor may simultaneously send data to the
centralized facility using a dial-up telephone line, and receives data from the centralized
facility through a signal encoded using a spread spectrum technique and modulated with a
broadcast signal from a third party.
WO 95/27349 describes apparatus and methods for including codes in audio signals and
decoding. An apparatus and methods for including a code having at least one code
frequency component in an audio signal are described. The abilities of various frequency
components in the audio signal to mask the code frequency component to human hearing
are evaluated, and based on these evaluations, an amplitude is assigned to the code
frequency components. Methods and apparatus for detecting a code in an encoded audio
signal are also described. A code frequency component in the encoded audio signal is
detected based on an expected code amplitude or on a noise amplitude within a range of
audio frequencies including the frequency of the code component.
However, when inserting the watermark information into a time/frequency spectrogram of
an audio signal, it is difficult to hide the watermark information below the masking
threshold or to find an optimal tradeoff between the assignment of as much energy as
possible to the watermark information - thus increasing the extractability at the decoder
side -, and keeping the watermark information being embedded inaudible when
reproducing the watermarked audio signal.
Summary of the Invention
In view of this situation, it is the object of the present invention to provide a scheme for
providing a watermark signal and a scheme for watermark embedding using that
watermark signal, which allows for a better trade-off between extractability and
inaudibility of the watermark signal.

This object is achieved by a watermark signal provider according to claim 1, a watermark
embedder according to claim 8, methods according to claim 9 or 10 and a computer
program according to claim 11.
According to an embodiment of the present invention, a watermark signal provider for
providing a watermark signal suitable for being hidden in an audio signal when the
watermark signal is added to the audio signal, such that the watermark signal represents
watermark data, comprises a psychoacoustical processor for determining a masking
threshold of the audio signal; and a modulator for generating the watermark signal from a
superposition of sample-shaping functions spaced apart from each other at a sample time
interval of a time-discrete representation of the watermark data, each sample-shaping
function being amplitude-weighted with a respective sample of the time-discrete
representation multiplied by a respective amplitude weight depending on the masking
threshold, the modulator being configured such that the sample time interval is shorter than
a time extension of the sample-shaping functions; and the respective amplitude weight also
depends on samples of the time-discrete representation neighboring the respective sample
in time.
The present invention is based on the finding that a better trade-off between extractability
and inaudibility of the watermark signal may be achieved by selecting the amplitude
weights for amplitude-weighting the sample-shaping functions which form, in
superposition, the watermarking signal, not only dependent on the masking threshold, but
also dependent on samples of the time-discrete representation of the watermark data
neighboring the respective sample. In this way, the sample-shaping functions at
neighboring sample positions may overlap each other, i.e. the sample time interval may be
shorter than the time extension of the sample-shaping function and, despite this,
interference between such neighboring sample-shaping functions may be compensated by
taking into account samples of the time-discrete representation neighboring the currently
weighted sample when setting the amplitude weight. Even further, since the sample-
shaping functions are allowed to have a larger time extension, their frequency responses
may be made narrower, thereby rendering the extractability of the watermark signal
stronger against reverberation, i.e. when the watermarked audio signal is reproduced in a
reverberant environment. In other words, the dependency of the respective amplitude
weight not only on the masking threshold, but also on samples of the time-discrete
representation of the watermark data neighboring the respective sample enables
compensating for audible interferences between neighboring sample-shaping functions,
which could otherwise lead to a violence of the masking threshold.

Brief Description of the Figures
Embodiments according to the invention will subsequently be described taking reference to
the enclosed figures, in which:
Fig. 1 shows a block schematic diagram of a watermark inserter according to an
embodiment of the invention;
Fig. 2 shows a block-schematic diagram of a watermark decoder, according to an
embodiment of the invention;
Fig. 3 shows a detailed block-schematic diagram of a watermark generator,
according to an embodiment of the invention;
Fig. 4 shows a detailed block-schematic diagram of a modulator, for use in an
embodiment of the invention;
Fig. 5 shows a detailed block-schematic diagram of a psychoacoustical processing
module, for use in an embodiment of the invention;
Fig. 6 shows a block-schematic diagram of a psychoacoustical model processor,
for use in an embodiment of the invention;
Fig. 7 shows a graphical representation of a power spectrum of an audio signal
output by block 801 over frequency;
Fig. 8 shows a graphical representation of a power spectrum of an audio signal
output by block 802 over frequency;
Fig. 9 shows a block-schematic diagram of an amplitude calculation;
Fig. 10a shows a block schematic diagram of a modulator;
Fig. 10b shows a graphical representation of the location of coefficients on the time-
frequency claim;
Figs. 11a and lib show a block-schematic diagrams of implementation alternatives of
the synchronization module;

Fig. 12a shows a graphical representation of the problem of finding the temporal
alignment of a watermark;
Fig. 12b shows a graphical representation of the problem of identifying the message
start;
Fig. 12c shows a graphical representation of a temporal alignment of synchronization
sequences in a full message synchronization mode;
Fig. 12d shows a graphical representation of the temporal alignment of the
synchronization sequences in a partial message synchronization mode;
Fig. 12e shows a graphical representation of input data of the synchronization
module;
Fig. 12f shows a graphical representation of a concept of identifying a
synchronization hit;
Fig. 12g shows a block-schematic diagram of a synchronization signature correlator;
Fig. 13a shows a graphical representation of an example for a temporal despreading;
Fig. 13b shows a graphical representation of an example for an element-wise
multiplication between bits and spreading sequences;
Fig. 13c shows a graphical representation of an output of the synchronization
signature correlator after temporal averaging;
Fig. 13d shows a graphical representation of an output of the synchronization
signature correlator filtered with the auto-correlation function of the
synchronization signature;
Fig. 14 shows a block-schematic diagram of a watermark extractor, according to an
embodiment of the invention;
Fig. 15 shows a schematic representation of a selection of a part of the time-
frequency-domain representation as a candidate message;

Fig. 16 shows a block-schematic diagram of an analysis module;
Fig. 17a shows a graphical representation of an output of a synchronization
correlator;
Fig. 17b shows a graphical representation of decoded messages;
Fig. 17c shows a graphical representation of a synchronization position, which is
extracted from a watermarked signal;
Fig. 18a shows a graphical representation of a payload, a payload with a Viterbi
termination sequence, a Viterbi-encoded payload and a repetition-coded
version of the Viterbi-coded payload;
Fig, 18b shows a graphical representation of subcarriers used for embedding a
watermarked signal;
Fig. 19 shows a graphical representation of an uncoded message, a coded message,
a synchronization message and a watermark signal, in which the
synchronization sequence is applied to the messages;
Fig. 20 shows a schematic representation of a first step of a so-called "ABC
synchronization" concept;
Fig. 21 shows a graphical representation of a second step of the so-called "ABC
synchronization" concept;
Fig. 22 shows a graphical representation of a third step of the so-called "ABC
synchronization" concept;
Fig. 23 shows a graphical representation of a message comprising a payload and a
CRC portion;
Fig. 24 shows a block-schematic diagram of a watermark signal provider according
to an embodiment of the invention; and

Fig. 25 shows a block-schematic diagram of a watermark embedder according to an
embodiment of the present invention.
Detailed Description of the Embodiments
1. Watermark signal provision
In the following, a watermark signal provider 2400 will be described referring to Fig. 24.
The watermark signal provider 2400 comprises a psychoacoustical processor 2410 and a
modulator 2420. The psychoacoustical processor 2410 is configured to receive the audio
signal 2430 for which the watermark signal provider 2400 is to provide the watermark
signal 2440. The modulator 2420, in turn, is configured to use the masking threshold
provided by the psychoacoustical processor 2410 in order to generate the watermark signal
2440. In particular, modulator 2420 is configured to generate the watermark signal 2440
from a superposition of sample-shaping functions spaced apart from each other at a sample
time interval of a time-discreet representation of watermark data 2450 to be represented by
the watermark signal 2440. In particular, modulator 2420 uses the masking threshold when
generating the watermark signal 2440 such that the watermark signal 2440 is suitable for
being hidden in the audio signal 2430 when the watermark signal 2440 is added to the
audio signal 2430 in order to obtain a watermarked audio signal.
As is described in more detail below, the time-discrete representation of the watermark
data may, in fact, be a time/frequency-discrete representation and may be derived from the
watermark data 2450 by use of spreading in time domain and/or frequency domain. The
time or time/frequency grid to the grid positions of which the samples of the time-discrete
representation are assigned may be fixed in time and, especially, independent from the
audio signal 2430. The superposition, in turn, may be interpreted as a convolution of the
time/discrete representation having its samples arranged at the grid positions of the just-
mentioned grid, the samples being weighted with amplitude-weights which, in turn, not
only depend on the masking threshold but also on the samples of the time-discrete
representation neighboring in time.
The dependency of the amplitude-weights from the masking threshold may be as follows: a
amplitude-weight which is for being multiplied with a certain sample of the time-discrete
representation at a certain time block, is derived from the respective time block of the
masking threshold which, in turn, is itself time and frequency dependent. Thus, in case of a
time/frequency-discrete representation of the watermark data, each sample is multiplied

with a amplitude-weight which corresponds to masking threshold sampled at the respective
time/frequency grid position of that watermark representation sample.
Furthermore, it is possible to use time-differential coding for retrieving the time-discrete
representation from the watermark data 2450. Details on a specific embodiment are
describe below.
The modulator 2420 is configured to generate the watermark signal 2440 from the
superposition of the sample-shaping functions such that each sample-shaping function is
amplitude-weighted with a respective sample of the time-discrete representation multiplied
by a respective amplitude weight depending on the masking threshold determined by the
psychoacoustical processor 2410. In particular, modulator 2420 is configured such that the
sample time interval is shorter than a time extension of the sample-shaping function, and
such that the respective amplitude weight also depends on samples of the time-discrete
representation neighboring the respective sample.
As will be outlined in more detail below, the fact that the sample time interval is shorter
than the time extension of the sample-shaping functions results in an interference between
the sample-shaping functions neighboring in time, thereby increasing the risk of violating
the masking threshold by accident. Such a violence of the masking threshold is, however,
compensated for by making the amplitude weights also dependent on the samples of the
time-discrete representation neighboring the current sample.
In the embodiment for a watermark system outlined below, the just-mentioned dependency
is realized by an iterative setting of the amplitude weights. In particular, the
psychoacoustical processor 2410 may determine the masking threshold independent from
the watermark data, while the modulator 2420 may be configured to iteratively set the
amplitude weights by preliminarily determining the amplitude weights based on the
masking threshold independent from the watermark data. Modulator 2420 may then be
configured to check as to whether the superposition of the sample-shaping functions as
amplitude-weighted with the samples of the watermark representation multiplied by the
preliminarily-determined amplitude weights violates the masking threshold. If so, the
modulator 2420 may vary the preliminarily-determined amplitude weights so as to obtain a
further superposition. Modulator 2420 may repeat these iterations comprising the check
and the variation with the subsequent superposition until a respective break condition is
fulfilled such as the amplitude-weights maintaining their values within a certain variance
threshold. Since, in the above-mentioned check, the neighboring samples of the time-
discrete representation influence/interfere with each other due to the superposition and the

time extension of the sample-shaping functions exceeding the sample time interval, the
hole iterative process for generating is dependent on these neighboring samples of the
watermark data representation.
It should be noted that in the embodiments outlined below, a spreading of the watermark
data in time-domain is used in order to reveal the time-discrete representation just-
mentioned. However, such a time-spreading may be left away. The same applies to the
frequency-spreading also used in the embodiments outlined below.
2. Watermark embedder
Fig. 25 shows a watermark embedder using the watermark signal provider 2400 of Fig. 24.
In particular, the watermark embedder of Fig. 25 is generally indicated with the reference
number 2500 and comprises, besides the watermark signal provider 2400, an adder 2510
for adding the watermark signal 2440 as output by watermark signal provider 2400 and the
audio signal 2430 so as to obtain the watermarked audio signal 2530.
3. System Description
In the following, a system for a watermark transmission will be described, which
comprises a watermark inserter and a watermark decoder. Naturally, the watermark
inserter and the watermark decoder can be used independent from each other.
For the description of the system a top-down approach is chosen here. First, it is
distinguished between encoder and decoder. Then, in sections 3.1 to 3.5 each processing
block is described in detail.
The basic structure of the system can be seen in Figures 1 and 2, which depict the encoder
and decoder side, respectively. Fig 1 shows a block schematic diagram of a watermark
inserter 100. At the encoder side, the watermark signal 101b is generated in the processing
block 101 (also designated as watermark generator) from binary data 101a and on the basis
of information 104, 105 exchanged with the psychoacoustical processing module 102. The
information provided from block 102 typically guarantees that the watermark is inaudible.
The watermark generated by the watermark generatorlOl is then added to the audio signal
106. The watermarked signal 107 can then be transmitted, stored, or further processed. In
case of a multimedia file, e.g., an audio-video file, a proper delay needs to be added to the
video stream not to lose audio-video synchronicity. In case of a multichannel audio signal,

each channel is processed separately as explained in this document. The processing blocks
101 (watermark generator) and 102 (psychoacoustical processing module) are explained in
detail in Sections 3.1 and 3.2, respectively.
The decoder side is depicted in Figure 2, which shows a block schematic diagram of a
watermark detector 200. A watermarked audio signal 200a, e.g., recorded by a
microphone, is made available to the system 200. A first block 203, which is also
designated as an analysis module, demodulates and transforms the data (e.g., the
watermarked audio signal) in time/frequency domain (thereby obtaining a time-frequency-
domain representation 204 of the watermarked audio signal 200a) passing it to the
synchronization module 201, which analyzes the input signal 204 and carries out a
temporal synchronization, namely, determines the temporal alignment of the encoded data
(e.g. of the encoded watermark data relative to the time-frequency-domain representation).
This information (e.g., the resulting synchronization information 205) is given to the
watermark extractor 202, which decodes the data (and consequently provides the binary
data 202a, which represent the data content of the watermarked audio signal 200a).
3.1 The Watermark Generator 101
The watermark generator 101 is depicted detail in Figure 3. Binary data (expressed as ±1)
to be hidden in the audio signal 106 is given to the watermark generator 101. The block
301 organizes the data 101a in packets of equal length Mp. Overhead bits are added (e.g.
appended) for signaling purposes to each packet. Let Ms denote their number. Their use
will be explained in detail in Section 3.5. Note that in the following each packet of payload
bits together with the signaling overhead bits is denoted message.
Each message 301a, of length Nm = Ms + Mp, is handed over to the processing block 302,
the channel encoder, which is responsible of coding the bits for protection against errors. A
possible embodiment of this module consists of a convolutional encoder together with an
interleaver. The ratio of the convolutional encoder influences greatly the overall degree of
protection against errors of the watermarking system. The interleaver, on the other hand,
brings protection against noise bursts. The range of operation of the interleaver can be
limited to one message but it could also be extended to more messages. Let Rc denote the
code ratio, e.g., 1/4. The number of coded bits for each message is Nm/Rc. The channel
encoder provides, for example, an encoded binary message 302a.

The next processing block, 303, carries out a spreading in frequency domain. In order to
achieve sufficient signal to noise ratio, the information (e.g. the information of the binary
message 302a) is spread and transmitted in Nf carefully chosen subbands. Their exact
position in frequency is decided a priori and is known to both the encoder and the decoder.
Details on the choice of this important system parameter is given in Section 3.2.2. The
spreading in frequency is determined by the spreading sequence cf of size Nf X1. The
output 303a of the block 303 consists of Nf bit streams, one for each subband. The i-th bit
stream is obtained by multiplying the input bit with the i-th component of spreading
sequence cf. The simplest spreading consists of copying the bit stream to each output
stream, namely use a spreading sequence of all ones.
Block 304, which is also designated as a synchronization scheme inserter, adds a
synchronization signal to the bit stream. A robust synchronization is important as the
decoder does not know the temporal alignment of neither bits nor the data structure, i.e.,
when each message starts. The synchronization signal consists of Ns sequences of Nf bits
each. The sequences are multiplied element wise and periodically to the bit stream (or bit
streams 303a). For instance, let a, b, and c, be the Ns = 3 synchronization sequences (also
designated as synchronization spreading sequences). Block 304 multiplies a to the first
spread bit, b to the second spread bit, and c to the third spread bit. For the following bits
the process is periodically iterated, namely, a to the fourth bit, b for the fifth bit and so on.
Accordingly, a combined information-synchronization information 304a is obtained. The
synchronization sequences (also designated as synchronization spread sequences) are
carefully chosen to minimize the risk of a false synchronization. More details are given in
Section 3.4. Also, it should be noted that a sequence a, b, c,... may be considered as a
sequence of synchronization spread sequences.
Block 305 carries out a spreading in time domain. Each spread bit at the input, namely a
vector of length Nf, is repeated in time domain Nt times. Similarly to the spreading in
frequency, we define a spreading sequence ct of size Nt x1. The i-th temporal repetition is
multiplied with the i-th component of ct.
The operations of blocks 302 to 305 can be put in mathematical terms as follows. Let m of
size 1 xNm=Rc be a coded message, output of 302. The output 303a (which may be
considered as a spread information representation R) of block 303 is


the output 304a of block 304, which may be considered as a combined information-
synchronization representation C, is

where ° denotes the Schur element-wise product and

The output 305a of 305 is

where o and T denote the Kronecker product and transpose, respectively. Please recall that
binary data is expressed as ±1.
Block 306 performs a differential encoding of the bits. This step gives the system
additional robustness against phase shifts due to movement or local oscillator mismatches.
More details on this matter are given in Section 3.3. If b(i; j) is the bit for the i-th
frequency band and j-th time block at the input of block 306, the output bit bdiff (i; j) is

At the beginning of the stream, that is for j = 0, bdiff (ij - 1) is set to 1.
Block 307 carries out the actual modulation, i.e., the generation of the watermark signal
waveform depending on the binary information 306a given at its input. A more detailed
schematics is given in Figure 4. Nf parallel inputs, 401 to 40Nf contain the bit streams for
the different subbands. Each bit of each subband stream is processed by a bit shaping block
(411 to 41Nf). The output of the bit shaping blocks are waveforms in time domain. The
waveform generated for the j-th time block and i-th subband, denoted by Si.j(t), on the basis
of the input bit bdiff(i, j)

is computed as follows

where y(i; j) is a weighting factor provided by the psycho acoustical processing unit 102, Tb
is the bit time interval, and gi(t) is the bit forming function for the i-th subband. The bit
forming function is obtained from a baseband function modulated in frequency
with a cosine

where fj is the center frequency of the i-th subband and the superscript T stands for
transmitter. The baseband functions can be different for each subband. If chosen identical,
a more efficient implementation at the decoder is possible. See Section 3.3 for more
details.
The bit shaping for each bit is repeated in an iterative process controlled by the
psychoacoustical processing module (102). Iterations are necessary to fine tune the weights
y(i, j) to assign as much energy as possible to the watermark while keeping it inaudible.
More details are given in Section 3.2.
The complete waveform at the output of the i-th bit shaping fillter 41i is

The bit forming baseband function &i (*) is normally non zero for a time interval much
larger than Tb, although the main energy is concentrated within the bit interval. An
example can be seen if Figure 12a where the same bit forming baseband function is plotted
for two adjacent bits. In the figure we have Tb = 40 ms. The choice of Tb as well as the
shape of the function affect the system considerably. In fact, longer symbols provide
narrower frequency responses. This is particularly beneficial in reverberant environments.
In fact, in such scenarios the watermarked signal reaches the microphone via several
propagation paths, each characterized by a different propagation time. The resulting
channel exhibits strong frequency selectivity. Interpreted in time domain, longer symbols

are beneficial as echoes with a delay comparable to the bit interval yield constructive
interference, meaning that they increase the received signal energy. Notwithstanding,
longer symbols bring also a few drawbacks; larger overlaps might lead to intersymbol
interference (ISI) and are for sure more difficult to hide in the audio signal, so that the
psychoacoustical processing module would allow less energy than for shorter symbols.
The watermark signal is obtained by summing all outputs of the bit shaping filters

3.2 The Psychoacoustical Processing Module 102
As depicted in Figure 5, the psychoacoustical processing module 102 consists of 3 parts.
The first step is an analysis module 501 which transforms the time audio signal into the
time/frequency domain. This analysis module may carry out parallel analyses in different
time/frequency resolutions. After the analysis module, the time/frequency data is
transferred to the psychoacoustic model (PAM) 502, in which masking thresholds for the
watermark signal are calculated according to psychoacoustical considerations (see E.
Zwicker H.Fastl, "Psychoacoustics Facts and models"). The masking thresholds indicate
the amount of energy which can be hidden in the audio signal for each subband and time
block. The last block in the psychoacoustical processing module 102 depicts the amplitude
calculation module 503. This module determines the amplitude gains to be used in the
generation of the watermark signal so that the masking thresholds are satisfied, i.e., the
embedded energy is less or equal to the energy defined by the masking thresholds.
3.2.1 The Time/Frequency Analysis 501
Block 501 carries out the time/frequency transformation of the audio signal by means of a
lapped transform. The best audio quality can be achieved when multiple time/frequency
resolutions are performed. One efficient embodiment of a lapped transform is the short
time Fourier transform (STFT), which is based on fast Fourier transforms (FFT) of
windowed time blocks. The length of the window determines the time/frequency
resolution, so that longer windows yield lower time and higher frequency resolutions,
while shorter windows vice versa. The shape of the window, on the other hand, among
other things, determines the frequency leakage.

For the proposed system, we achieve an inaudible watermark by analyzing the data with
two different resolutions. A first filter bank is characterized by a hop size of Tt>, i.e., the bit
length. The hop size is the time interval between two adjacent time blocks. The window
length is approximately Tt,. Please note that the window shape does not have to be the
same as the one used for the bit shaping, and in general should model the human hearing
system. Numerous publications study this problem.
The second filter bank applies a shorter window. The higher temporal resolution achieved
is particularly important when embedding a watermark in speech, as its temporal structure
is in general finer than Tt,.
The sampling rate of the input audio signal is not important, as long as it is large enough to
describe the watermark signal without aliasing. For instance, if the largest frequency
component contained in the watermark signal is 6 kHz, then the sampling rate of the time
signals must be at least 12 kHz.
3.2.2 The Psychoacoustical Model 502
The psychoacoustical model 502 has the task to determine the masking thresholds, i.e., the
amount of energy which can be hidden in the audio signal for each subband and time block
keeping the watermarked audio signal indistinguishable from the original.
The i-th subband is defined between two limits, namely The subbands are
determined by defining Nf center frequencies fj and letting i for i = 2, 3, ... ,
Nf. An appropriate choice for the center frequencies is given by the Bark scale proposed
by Zwicker in 1961. The subbands become larger for higher center frequencies. A possible
implementation of the system uses 9 subbands ranging from 1.5 to 6 kHz arranged in an
appropriate way.
The following processing steps are carried out separately for each time/frequency
resolution for each subband and each time block. The processing step 801 carries out a
spectral smoothing. In fact, tonal elements, as well as notches in the power spectrum need
to be smoothed. This can be carried out in several ways. A tonality measure may be
computed and then used to drive an adaptive smoothing filter. Alternatively, in a simpler
implementation of this block, a median-like filter can be used. The median filter considers

a vector of values and outputs their median value. In a median-like filter the value
corresponding to a different quantile than 50% can be chosen. The filter width is defined in
Hz and is applied as a non-linear moving average which starts at the lower frequencies and
ends up at the highest possible frequency. The operation of 801 is illustrated in Figure 7.
The red curve is the output of the smoothing.
Once the smoothing has been carried out, the thresholds are computed by block 802
considering only frequency masking. Also in this case there are different possibilities. One
way is to use the minimum for each subband to compute the masking energy Ej. This is the
equivalent energy of the signal which effectively operates a masking. From this value we
can simply multiply a certain scaling factor to obtain the masked energy Jj. These factors
are different for each subband and time/frequency resolution and are obtained via empirical
psychoacoustical experiments. These steps are illustrated in Figure 8.
In block 805, temporal masking is considered. In this case, different time blocks for the
same subband are analyzed. The masked energies Jj are modified according to an
empirically derived postmasking profile. Let us consider two adjacent time blocks, namely
k-1 and k. The corresponding masked energies are Ji(k-l) and Jj(k). The postmasking
profile defines that, e.g., the masking energy Ej can mask an energy Jj at time k and a • Jj at
time k+1. In this case, block 805 compares Ji(k) (the energy masked by the current time
block) and a-Ji(k+l) (the energy masked by the previous time block) and chooses the
maximum. Postmasking profiles are available in the literature and have been obtained via
empirical psychoacoustical experiments. Note that for large Tb, i.e., > 20 ms, postmasking
is applied only to the time/frequency resolution with shorter time windows.
Summarizing, at the output of block 805 we have the masking thresholds per each subband
and time block obtained for two different time/frequency resolutions. The thresholds have
been obtained by considering both frequency and time masking phenomena. In block 806,
the thresholds for the different time/frequency resolutions are merged. For instance, a
possible implementation is that 806 considers all thresholds corresponding to the time and
frequency intervals in which a bit is allocated, and chooses the minimum.
3.2.3 The Amplitude Calculation Block 503
Please refer to Figure 9. The input of 503 are the thresholds 505 from the psychoacoustical
model 502 where all psycho acoustics motivated calculations are carried out. In the
amplitude calculator 503 additional computations with the thresholds are performed. First,
an amplitude mapping 901 takes place. This block merely converts the masking thresholds

(normally expressed as energies) into amplitudes which can be used to scale the bit shaping
function defined in Section 3.1. Afterwards, the amplitude adaptation block 902 is run.
This block iteratively adapts the amplitudes y(i, j) which are used to multiply the bit
shaping functions in the watermark generator 101 so that the masking thresholds are
indeed fulfilled. In fact, as already discussed, the bit shaping function normally extends for
a time interval larger than Tb Therefore, multiplying the correct amplitude y(i, j) which
fulfills the masking threshold at point i, j does not necessarily fulfill the requirements at
point i, j-1. This is particularly crucial at strong onsets, as a preecho becomes audible.
Another situation which needs to be avoided is the unfortunate superposition of the tails of
different bits which might lead to an audible watermark. Therefore, block 902 analyzes the
signal generated by the watermark generator to check whether the thresholds have been
fulfilled. If not, it modifies the amplitudes y(i, j) accordingly.
This concludes the encoder side. The following sections deal with the processing steps
carried out at the receiver (also designated as watermark decoder).
3.3 The Analysis Module 203
The analysis module 203 is the first step (or block) of the watermark extraction process. Its
purpose is to transform the watermarked audio signal 200a back into Nf bit streams b, (j)
(also designated with 204), one for each spectral subband i. These are further processed by
the synchronization module 201 and the watermark extractor 202, as discussed in Sections
3.4 and 3.5, respectively. Note that the bt(j) are soft bit streams, i.e., they can take, for
example, any real value and no hard decision on the bit is made yet.
The analysis module consists of three parts which are depicted in Figure 16: The analysis
filter bank 1600, the amplitude normalization block 1604 and the differential decoding
1608.
3.3.1 Analysis filter bank 1600
The watermarked audio signal is transformed into the time-frequency domain by the
analysis filter bank 1600 which is shown in detail in Figure 10a. The input of the filter
bank is the received watermarked audio signal r(t). Its output are the complex coefficients

for the i-th branch or subband at time instant j. These values contain information
about the amplitude and the phase of the signal at center frequency f; and time j-Tb.
The filter bank 1600 consists of Nf branches, one for each spectral subband i. Each branch
splits up into an upper subbranch for the in-phase component and a lower subbranch for
the quadrature component of the subband i. Although the modulation at the watermark
generator and thus the watermarked audio signal are purely real-valued, the complex-
valued analysis of the signal at the receiver is needed because rotations of the modulation
constellation introduced by the channel and by synchronization misalignments are not
known at the receiver. In the following we consider the i-th branch of the filter bank. By
combining the in-phase and the quadrature subbranch, we can define the complex-valued
baseband signal

where * indicates convolution and is the impulse response of the receiver lowpass
filter of subband i. Usually J is equal to the baseband bit forming function ' of
subband i in the modulator 307 in order to fulfill the matched filter condition, but other
impulse responses are possible as well.
In order to obtain the coefficients with rate l=Tb, the continuous output
must be sampled. If the correct timing of the bits was known by the receiver, sampling
with rate l=Tb would be sufficient. However, as the bit synchronization is not known yet,
sampling is carried out with rate Nos/Tb where Nos is the analysis filter bank oversampling
factor. By choosing Nos sufficiently large (e.g. Nos = 4), we can assure that at least one
sampling cycle is close enough to the ideal bit synchronization. The decision on the best
oversampling layer is made during the synchronization process, so all the oversampled
data is kept until then. This process is described in detail in Section 3.4.
At the output of the i-th branch we have the coefficients where j indicates the bit
number or time instant and k indicates the oversampling position within this single bit,
where k= 1;2; ....,N0S.
Figure 10b gives an exemplary overview of the location of the coefficients on the time-
frequency plane. The oversampling factor is Nos = 2. The height and the width of the

rectangles indicate respectively the bandwidth and the time interval of the part of the signal
that is represented by the corresponding coefficient
If the subband frequencies fj are chosen as multiples of a certain interval Af the analysis
filter bank can be efficiently implemented using the Fast Fourier Transform (FFT).
3.3.2 Amplitude normalization 1604
Without loss of generality and to simplify the description, we assume that the bit
synchronization is known and that in the following. That is, we have complex
coeffcients at the input of the normalization block 1604. As no channel state
information is available at the receiver (i.e., the propagation channel in unknown), an equal
gain combining (EGC) scheme is used. Due to the time and frequency dispersive channel,
the energy of the sent bit bi(j) is not only found around the center frequency fj and time
instant j, but also at adjacent frequencies and time instants. Therefore, for a more precise
weighting, additional coefficients at frequencies are calculated and used for
normalization of coefficient we have, for example,

The normalization for n > 1 is a straightforward extension of the formula above. In the
same fashion we can also choose to normalize the soft bits by considering more than one
time instant. The normalization is carried out for each subband i and each time instant j.
The actual combining of the EGC is done at later steps of the extraction process.
3.3.3 Differential decoding 1608
At the input of the differential decoding block 1608 we have amplitude normalized
complex coefficients(j)Hvhich contain information about the phase of the signal
components at frequency fj and time instant j. As the bits are differentially encoded at the
transmitter, the inverse operation must be performed here. The soft bits are obtained
by first calculating the difference in phase of two consecutive coefficients and then taking
the real part:


This has to be carried out separately for each subband because the channel normally
introduces different phase rotations in each subband.
3.4 The Synchronization Module 201
The synchronization module's task is to find the temporal alignment of the watermark. The
problem of synchronizing the decoder to the encoded data is twofold. In a first step, the
analysis filterbank must be aligned with the encoded data, namely the bit shaping functions
9i (f-) used in the synthesis in the modulator must be aligned with the filters 9'i '(*) used
for the analysis. This problem is illustrated in Figure 12a, where the analysis filters are
identical to the synthesis ones. At the top, three bits are visible. For simplicity, the
waveforms for all three bits are not scaled. The temporal offset between different bits is Tb.
The bottom part illustrates the synchronization issue at the decoder: the filter can be
applied at different time instants, however, only the position marked in red (curve 1299a)
is correct and allows to extract the first bit with the best signal to noise ratio SNR and
signal to interference ratio SIR. In fact, an incorrect alignment would lead to a degradation
of both SNR and SIR. We refer to this first alignment issue as "bit synchronization". Once
the bit synchronization has been achieved, bits can be extracted optimally. However, to
correctly decode a message, it is necessary to know at which bit a new message starts. This
issue is illustrated in Figure 12b and is referred to as message synchronization. In the
stream of decoded bits only the starting position marked in red (position 1299b) is correct
and allows to decode the k-th message.
We first address the message synchronization only. The synchronization signature, as
explained in Section 3.1, is composed of Ns sequences in a predetermined order which are
embedded continuously and periodically in the watermark. The synchronization module is
capable of retrieving the temporal alignment of the synchronization sequences. Depending

on the size Ns we can distinguish between two modes of operation, which are depicted in
Figure 12c and 12d, respectively.
In the full message synchronization mode (Fig. 12c) we have Ns = Nm/Rc. For simplicity in
the figure we assume Ns = Nm/Rc = 6 and no time spreading, i.e., Nt = 1. The
synchronization signature used, for illustration purposes, is shown beneath the messages.
In reality, they are modulated depending on the coded bits and frequency spreading
sequences, as explained in Section 3.1. In this mode, the periodicity of the synchronization
signature is identical to the one of the messages. The synchronization module therefore can
identify the beginning of each message by finding the temporal alignment of the
synchronization signature. We refer to the temporal positions at which a new
synchronization signature starts as synchronization hits. The synchronization hits are then
passed to the watermark extractor 202.
The second possible mode, the partial message synchronization mode (Fig. 12d), is
depicted in Figure 12d. In this case we have Ns < Nm=Rc. In the figure we have taken Ns =
3, so that the three synchronization sequences are repeated twice for each message. Please
note that the periodicity of the messages does not have to be multiple of the periodicity of
the synchronization signature. In this mode of operation, not all synchronization hits
correspond to the beginning of a message. The synchronization module has no means of
distinguishing between hits and this task is given to the watermark extractor 202.
The processing blocks of the synchronization module are depicted in Figures 1 la and 1 lb.
The synchronization module carries out the bit synchronization and the message
synchronization (either full or partial) at once by analyzing the output of the
synchronization signature correlator 1201. The data in time/frequency domain 204 is
provided by the analysis module. As the bit synchronization is not yet available, block 203
oversamples the data with factor Nos, as described in Section 3.3. An illustration of the
input data is given in Figure 12e. For this example we have taken Nos = 4, Nt = 2, and Ns =
3. In other words, the synchronization signature consists of 3 sequences (denoted with a, b,
and c). The time spreading, in this case with spreading sequence ct = [1 1] T, simply repeats
each bit twice in time domain. The exact synchronization hits are denoted with arrows and
correspond to the beginning of each synchronization signature. The period of the
synchronization signature is Nt • Nos ■ Ns = Nsbi which is 2 * 4 • 3 = 24, for example. Due to
the periodicity of the synchronization signature, the synchronization signature correlator
(1201) arbitrarily divides the time axis in blocks, called search blocks, of size Nsbi, whose
subscript stands for search block length. Every search block must contain (or typically
contains) one synchronization hit as depicted in Figure 12f. Each of the Nst>i bits is a

candidate synchronization hit. Block 1201's task is to compute a likelihood measure for
each of candidate bit of each block. This information is then passed to block 1204 which
computes the synchronization hits.
3.4.1 The synchronization signature correlator 1201
For each of the NSbi candidate synchronization positions the synchronization signature
correlator computes a likelihood measure, the latter is larger the more probable it is that the
temporal alignment (both bit and partial or full message synchronization) has been found.
The processing steps are depicted in Figure 12g.
Accordingly, a sequence 1201aof likelihood values, associated with different positional
choices, may be obtained.
Block 1301 carries out the temporal despreading, i.e., multiplies every Nt bits with the
temporal spreading sequence ct and then sums them. This is carried out for each of the Nf
frequency subbands. Figure 13a shows an example. We take the same parameters as
described in the previous section, namely Nos = 4, Nt = 2, and Ns = 3. The candidate
synchronization position is marked. From that bit, with Nos offset, Nt • Ns are taken by
block 1301 and time despread with sequence ct, so thatNs bits are left.
In block 1302 the bits are multiplied element-wise with the Ns spreading sequences (see
Figure 13b).
In block 1303 the frequency despreading is carried out, namely, each bit is multiplied with
the spreading sequence Cf and then summed along frequency.
At this point, if the synchronization position were correct, we would have Ns decoded bits.
As the bits are not known to the receiver, block 1304 computes the likelihood measure by
taking the absolute values of the Ns values and sums.
The output of block 1304 is in principle a non coherent correlator which looks for the
synchronization signature. In fact, when choosing a small Ns, namely the partial message
synchronization mode, it is possible to use synchronization sequences (e.g. a, b, c) which
are mutually orthogonal. In doing so, when the correlator is not correctly aligned with the
signature, its output will be very small, ideally zero. When using the full message
synchronization mode it is advised to use as many orthogonal synchronization sequences

as possible, and then create a signature by carefully choosing the order in which they are
used. In this case, the same theory can be applied as when looking for spreading sequences
with good auto correlation functions. When the correlator is only slightly misaligned, then
the output of the correlator will not be zero even in the ideal case, but anyway will be
smaller compared to the perfect alignment, as the analysis filters cannot capture the signal
energy optimally.
3.4.2 Synchronization hits computation 1204
This block analyzes the output of the synchronization signature correlator to decide where
the synchronization positions are. Since the system is fairly robust against misalignments
of up to TV4 and the Tb is normally taken around 40 ms, it is possible to integrate the
output of 1201 over time to achieve a more stable synchronization. A possible
implementation of this is given by an IIR filter applied along time with a exponentially
decaying impulse response. Alternatively, a traditional FIR moving average filter can be
applied. Once the averaging has been carried out, a second correlation along different Nt-Ns
is carried out ("different positional choice")- In fact, we want to exploit the information
that the autocorrelation function of the synchronization function is known. This
corresponds to a Maximum Likelihood estimator. The idea is shown in Figure 13c. The
curve shows the output of block 1201 after temporal integration. One possibility to
determine the synchronization hit is simply to find the maximum of this function. In Figure
13d we see the same function (in black) filtered with the autocorrelation function of the
synchronization signature. The resulting function is plotted in red. In this case the
maximum is more pronounced and gives us the position of the synchronization hit. The
two methods are fairly similar for high SNR but the second method performs much better
in lower SNR regimes. Once the synchronization hits have been found, they are passed to
the watermark extractor 202 which decodes the data.
In some embodiments, in order to obtain a robust synchronization signal, synchronization
is performed in partial message synchronization mode with short synchronization
signatures. For this reason many decodings have to be done, increasing the risk of false
positive message detections. To prevent this, in some embodiments signaling sequences
may be inserted into the messages with a lower bit rate as a consequence.
This approach is a solution to the problem arising from a sync signature shorter than the
message, which is already addressed in the above discussion of the enhanced
synchronization. In this case, the decoder doesn't know where a new message starts and

attempts to decode at several synchronization points. To distinguish between legitimate
messages and false positives, in some embodiments a signaling word is used (i.e. payload
is sacrified to embed a known control sequence). In some embodiments, a plausibility
check is used (alternatively or in addition) to distinguish between legitimate messages and
false positives.
3.5 The Watermark Extractor 202
The parts constituting the watermark extractor 202 are depicted in Figure 14. This has two
inputs, namely 204 and 205 from blocks 203 and 201, respectively. The synchronization
module 201 (see Section 3.4) provides synchronization timestamps, i.e., the positions in
time domain at which a candidate message starts. More details on this matter are given in
Section 3.4. The analysis interbank block 203, on the other hand, provides the data in
time/frequency domain ready to be decoded.
The first processing step, the data selection block 1501, selects from the input 204 the part
identified as a candidate message to be decoded. Figure 15 shows this procedure
graphically. The input 204 consists of Nf streams of real values. Since the time alignment is
not known to the decoder a priori, the analysis block 203 carries out a frequency analysis
with a rate higher than 1/Tb Hz (oversampling). In Figure 15 we have used an
oversampling factor of 4, namely, 4 vectors of size NfX 1 are output every Tb seconds.
When the synchronization block 201 identifies a candidate message, it delivers a
timestamp 205 indicating the starting point of a candidate message. The selection block
1501 selects the information required for the decoding, namely a matrix of size Nf x Nm/Rc.
This matrix 1501a is given to block 1502 for further processing.
Blocks 1502, 1503, and 1504 carry out the same operations of blocks 1301, 1302, and
1303 explained in Section 3.4.
An alternative embodiment of the invention consists in avoiding the computations done in
1502-1504 by letting the synchronization module deliver also the data to be decoded.
Conceptually it is a detail. From the implementation point of view, it is just a matter of
how the buffers are realized. In general, redoing the computations allows us to have
smaller buffers.
The channel decoder 1505 carries out the inverse operation of block 302. If channel
encoder, in a possible embodiment of this module, consisted of a convolutional encoder
together with an interleaver, then the channel decoder would perform the deinterleaving

and the convolutional decoding, e.g., with the well known Viterbi algorithm. At the output
of this block we have Nm bits, i.e., a candidate message.
Block 1506, the signaling and plausibility block, decides whether the input candidate
message is indeed a message or not. To do so, different strategies are possible.
The basic idea is to use a signaling word (like a CRC sequence) to distinguish between true
and false messages. This however reduces the number of bits available as payload.
Alternatively we can use plausibility checks. If the messages for instance contain a
timestamp, consecutive messages must have consecutive timestamps. If a decoded message
possesses a timestamp which is not the correct order, we can discard it.
When a message has been correctly detected the system may choose to apply the look
ahead and/or look back mechanisms. We assume that both bit and message
synchronization have been achieved. Assuming that the user is not zapping, the system
"looks back" in time and attempts to decode the past messages (if not decoded already)
using the same synchronization point (look back approach). This is particularly useful
when the system starts. Moreover, in bad conditions, it might take 2 messages to achieve
synchronization. In this case, the first message has no chance. With the look back option
we can save "good" messages which have not been received only due to back
synchronization. The look ahead is the same but works in the future. If we have a message
now we know where the next message should be, and we can attempt to decode it anyhow.
3.6. Synchronization Details
For the encoding of a payload, for example, a Viterbi algorithm may be used. Fig. 18a
shows a graphical representation of a payload 1810, a Viterbi termination sequence 1820, a
Viterbi encoded payload 1830 and a repetition-coded version 1840 of the Viterbi-coded
payload. For example, the payload length may be 34 bits and the Viterbi termination
sequence may comprise 6 bits. If, for example a Viterbi code rate of 1/7 may be used the
Viterbi-coded payload may comprise (34+6)*7=280 bits. Further, by using a repetition
coding of 1/2, the repetition coded version 1840 of the Viterbi-encoded payload 1830 may
comprise 280*2=560 bits. In this example, considering a bit time interval of 42.66 ms, the
message length would be 23.9 s. The signal may be embedded with, for example, 9
subcarriers (e.g. placed according to the critical bands) from 1.5 to 6 kHz as indicated by
the frequency spectrum shown in Fig. 18b. Alternatively, also another number of

subcarriers (e.g. 4, 6, 12, 15 or a number between 2 and 20) within a frequency range
between 0 and 20 kHz maybe used.
Fig. 19 shows a schematic illustration of the basic concept 1900 for the synchronization,
also called ABC synch. It shows a schematic illustration of an uncoded messages 1910, a
coded message 1920 and a synchronization sequence (synch sequence) 1930 as well as the
application of the synch to several messages 1920 following each other.
The synchronization sequence or synch sequence mentioned in connection with the
explanation of this synchronization concept (shown in Fig. 19 - 23) may be equal to the
synchronization signature mentioned before.
Further, Fig. 20 shows a schematic illustration of the synchronization found by correlating
with the synch sequence. If the synchronization sequence 1930 is shorter than the message,
more than one synchronization point 1940 (or alignment time block) may be found within
a single message. In the example shown in Fig. 20, 4 synchronization points are found
within each message. Therefore, for each synchronization found, a Viterbi decoder (a
Viterbi decoding sequence) may be started. In this way, for each synchronization point
1940 a message 2110 may be obtained, as indicated in Fig. 21.
Based on these messages the true messages 2210 may be identified by means of a CRC
sequence (cyclic redundancy check sequence) and/or a plausibility check, as shown in Fig.
22.
The CRC detection (cyclic redundancy check detection) may use a known sequence to
identify true messages from false positive. Fig. 23 shows an example for a CRC sequence
added to the end of a payload.
The probability of false positive (a message generated based on a wrong synchronization
point) may depend on the length of the CRC sequence and the number of Viterbi decoders
(number of synchronization points within a single message) started. To increase the length
of the payload without increasing the probability of false positive a plausibility may be
exploited (plausibility test) or the length of the synchronization sequence (synchronization
signature) may be increased.
4. Concepts and Advantages

In the following, some aspects of the above discussed system will be described, which are
considered as being innovative. Also, the relation of those aspects to the state-of-the-art
technologies will be discussed.
4.1. Continuous synchronization
Some embodiments allow for a continuous synchronization. The synchronization signal,
which we denote as synchronization signature, is embedded continuously and parallel to
the data via multiplication with sequences (also designated as synchronization spread
sequences) known to both transmit and receive side.
Some conventional systems use special symbols (other than the ones used for the data),
while some embodiments according to the invention do not use such special symbols.
Other classical methods consist of embedding a known sequence of bits (preamble) time-
multiplexed with the data, or embedding a signal frequency-multiplexed with the data.
However, it has been found that using dedicated sub-bands for synchronization is
undesired, as the channel might have notches at those frequencies, making the
synchronization unreliable. Compared to the other methods, in which a preamble or a
special symbol is time-multiplexed with the data, the method described herein is more
advantageous as the method described herein allows to track changes in the
synchronization (due e.g. to movement) continuously.
Furthermore, the energy of the watermark signal is unchanged (e.g. by the multiplicative
introduction of the watermark into the spread information representation), and the
synchronization can be designed independent from the psychoacoustical model and data
rate. The length in time of the synchronization signature, which determines the robustness
of the synchronization, can be designed at will completely independent of the data rate.
Another classical method consists of embedding a synchronization sequence code-
multiplexed with the data. When compared to this classical method, the advantage of the
method described herein is that the energy of the data does not represent an interfering
factor in the computation of the correlation, bringing more robustness. Furthermore, when
using code-multiplexing, the number of orthogonal sequences available for the
synchronization is reduced as some are necessary for the data.

To summarize, the continuous synchronization approach described herein brings along a
large number of advantages over the conventional concepts.
However, in some embodiments according to the invention, a different synchronization
concept may be applied.
4.2. 2D spreading
Some embodiments of the proposed system carry out spreading in both time and frequency
domain, i.e. a 2-dimensional spreading (briefly designated as 2D-spreading). It has been
found that this is advantageous with respect to ID systems as the bit error rate can be
further reduced by adding redundance in e.g. time domain.
However, in some embodiments according to the invention, a different spreading concept
may be applied.
4.3. Differential encoding and Differential decoding
In some embodiments according to the invention, an increased robustness against
movement and frequency mismatch of the local oscillators (when compared to
conventional systems) is brought by the differential modulation. It has been found that in
fact, the Doppler effect (movement) and frequency mismatches lead to a rotation of the
BPSK constellation (in other words, a rotation on the complex plane of the bits). In some
embodiments, the detrimental effects of such a rotation of the BPSK constellation (or any
other appropriate modulation constellation) are avoided by using a differential encoding or
differential decoding.
However, in some embodiments according to the invention, a different encoding concept
or decoding concept may be applied. Also, in some cases, the differential encoding may be
omitted.
4.4. Bit shaping

In some embodiments according to the invention, bit shaping brings along a significant
improvement of the system performance, because the reliability of the detection can be
increased using a filter adapted to the bit shaping.
In accordance with some embodiments, the usage of bit shaping with respect to
watermarking brings along improved reliability of the watermarking process. It has been
found that particularly good results can be obtained if the bit shaping function is longer
than the bit interval.
However, in some embodiments according to the invention, a different bit shaping concept
may be applied. Also, in some cases, the bit shaping may be omitted.
4.5. Interactive between Psychoacoustic Model (PAM) and Filter Bank (FB) synthesis
In some embodiments, the psychoacoustical model interacts with the modulator to fine
tune the amplitudes which multiply the bits.
However, in some other embodiments, this interaction may be omitted.
4.6. Look ahead and look back features
In some embodiments, so called "Look back" and "look ahead" approaches are applied.
In the following, these concepts will be briefly summarized. When a message is correctly
decoded, it is assumed that synchronization has been achieved. Assuming that the user is
not zapping, in some embodiments a look back in time is performed and it is tried to
decode the past messages (if not decoded already) using the same synchronization point
(look back approach). This is particularly useful when the system starts.
In bad conditions, it might take 2 messages to achieve synchronization. In this case, the
first message has no chance in conventional systems. With the look back option, which is
used in some embodiments of the invention, it is possible to save (or decode) "good"
messages which have not been received only due to back synchronization.
The look ahead is the same but works in the future. If I have a message now I know where
my next message should be, and I can try to decode it anyhow. Accordingly, overlapping
messages can be decoded.

However, in some embodiments according to the invention, the look ahead feature and/or
the look back feature may be omitted.
4.7. Increased synchronization robustness
In some embodiments, in order to obtain a robust synchronization signal, synchronization
is performed in partial message synchronization mode with short synchronization
signatures. For this reason many decodings have to be done, increasing the risk of false
positive message detections. To prevent this, in some embodiments signaling sequences
may be inserted into the messages with a lower bit rate as a consequence.
However, in some embodiments according to the invention, a different concept for
improving the synchronization robustness may be applied. Also, in some cases, the usage
of any concepts for increasing the synchronization robustness may be omitted.
4.8. Other enhancements
In the following, some other general enhancements of the above described system with
respect to background art will be put forward and discussed:
1. lower computational complexity
2. better audio quality due to the better psychoacoustical model
3. more robustness in reverberant environments due to the narrowband multicarrier
signals
4. an SNR estimation is avoided in some embodiments. This allows for better
robustness, especially in low SNR regimes.
Some embodiments according to the invention are better than conventional systems, which
use very narrow bandwidths of, for example, 8Hz for the following reasons:
1. 8 Hz bandwidths (or a similar very narrow bandwidth) requires very long time
symbols because the psychoacoustical model allows very little energy to make it inaudible;

2. 8 Hz (or a similar very narrow bandwidth) makes it sensitive against time varying
Doppler spectra. Accordingly, such a narrow band system is typically not good enough if
implemented, e.g., in a watch.
Some embodiments according to the invention are better than other technologies for the
following reasons:
1. Techniques which input an echo fail completely in reverberant rooms. In contrast,
in some embodiments of the invention, the introduction of an echo is avoided.
2. Techniques which use only time spreading have longer message duration in
comparison embodiments of the above described system in which a two-dimensional
spreading, for example both in time and in frequency, is used.
Some embodiments according to the invention are better than the system described in DE
196 40 814, because one of more of the following disadvantages of the system according to
said document are overcome:
• the complexity in the decoder according to DE 196 40 814 is very high, a filter of
length 2N with N = 128 is used
• the system according to DE 196 40 814 comprises a long message duration
• in the system according to DE 196 40 814 spreading only in time domain with
relatively high spreading gain (e.g. 128)
• in the system according to DE 196 40 814 the signal is generated in time domain,
transformed to spectral domain, weighted, transformed back to time domain, and
superposed to audio, which makes the system very complex
5. Applications
The invention comprises a method to modify an audio signal in order to hide digital data
and a corresponding decoder capable of retrieving this information while the perceived
quality of the modified audio signal remains indistinguishable to the one of the original.
Examples of possible applications of the invention are given in the following:

1. Broadcast monitoring: a watermark containing information on e.g. the station and
time is hidden in the audio signal of radio or television programs. Decoders, incorporated
in small devices worn by test subjects, are capable to retrieve the watermark, and thus
collect valuable information for advertisements agencies, namely who watched which
program and when.
2. Auditing: a watermark can be hidden in, e.g., advertisements. By automatically
monitoring the transmissions of a certain station it is then possible to know when exactly
the ad was broadcast. In a similar fashion it is possible to retrieve statistical information
about the programming schedules of different radios, for instance, how often a certain
music piece is played, etc.
3. Metadata embedding: the proposed method can be used to hide digital information
about the music piece or program, for instance the name and author of the piece or the
duration of the program etc.
Summarizing the above embodiments and comparing the embodiments of Fig. 1 to 23 with
the embodiments of Fig. 24 and 25, these embodiments described a watermark signal
provider 2400 for providing a watermark signal 2440; 101b suitable for being hidden in an
audio signal 2430; 106 when the watermark signal is added to the audio signal, such that
the watermark signal represents watermark data 2450; 101a, the watermark signal provider
comprising a psychoacoustical processor 2410; 102 for determining a masking threshold of
the audio signal; and a modulator 2420; 307 in 101 for generating the watermark signal
from a superposition as represented by equation 8 and shown in Fig. 12a, for example, of
sample-shaping functions Qi (*) spaced apart from each other at a sample time interval Tb
of a time-discrete representation bdjfr (i, j) of the watermark data, namely the above-
mentioned packets of equal length Mp, each sample-shaping function 0i"(*) being
amplitude-weighted with a respective sample bdiff (i, j) of the time-discrete representation,
multiplied by a respective amplitude weight y(i; j) depending on the masking threshold, the
modulator being configured such that the sample time interval Tb is shorter than a time
extension of the sample-shaping functions as exemplarily shown in Fig. 12a; and the
respective amplitude weight y(i; j) also depends on samples of the time-discrete
representation neighboring the respective sample in time.
In particular, the psychoacoustical processor may be configured to determine the masking
threshold independent from the watermark data 2450 and the modulator may be configured
to generate the watermark signal iteratively by preliminarily determining a preliminary

amplitude weight y(i; j) based on the masking threshold independent from the watermark
data, and then checking as to whether the superposition of the sample-shaping functions
using the preliminary amplitude weight as the respective amplitude weight violates the
masking threshold. If so, then the preliminary amplitude weight is varied so as to obtain a
superposition of the sample-shaping functions using the varied amplitude weight as the
respective amplitude weight. As already outlined above, since in the check, the
neighboring samples of the time-discrete representation influence/interfere with each other
due to the superposition and the time extension of the sample-shaping functions exceeding
the sample time interval, the hole iterative process for generating the watermark signal
2440 and the finally used amplitude-weightings, respectively,, are dependent on these
neighboring samples of the watermark data representation. In other words, the check
induces a dependency of the finally used amplitude weights 7(1; j) from the samples bdiff (i,
j±l) and enables a good tradeoff between watermark extractability and inaudibility of the
watermark signal. Of course, the procedure of checking, superpositioning and varying may
iteratively repeated.
The just-mentioned dependency on the neighboring samples of the watermark data
representation may, alternatively, be implemented by non-iteratively setting the amplitude-
weightings. For example, the modulator may analytically determine the amplitude weights
y(i; j) based on both the masking threshold at (i j) as well as the neighboring watermark
samples bdiff (i,j±l).
A time-spreader 305 may be used to spread the watermark data in time in order to obtain
the time-discrete representation. Further, a frequency-spreader 303 may be used to spread
the watermark data in a frequency domain in order to obtain the time-discrete
representation. A time/frequency analyzer 501 may be used transfer the audio signal from
the time domain to a frequency domain by means of a lapped transform using a first
window length of approximately the sample time interval. The time/frequency analyzer
may be configured to transfer the audio signal from the time domain to the frequency
domain by means of the lapped transform also using a second window length being shorter
than the first window length.
When the time-discrete representation is composed of time-discrete subbands, the
modulator may be configured to generate the watermark signal from, for each time-discrete
subband, a superposition according to both equation 8 and 9 of sample-shaping functions
spaced apart at the sample time interval with each sample-shaping function being
amplitude-weighted with a respective sample of the respective time-discrete subband
multiplied by a respective amplitude weight depending on the masking threshold, the

sample-shaping functions of the superposition for a respective time-discrete subband
comprising a carrier frequency at a center frequency f, of the respective time-discreet
subband i.
Further, the above embodiments described a watermark embedder 2500; 100 comprising a
watermark signal provider 2400 and an adder 2510 for adding the watermark signal and the
audio signal to obtain a watermarked audio signal.
6. Implementation Alternatives
Although some aspects have been described in the context of an apparatus, it is clear that
these aspects also represent a description of the corresponding method, where a block or
device corresponds to a method step or a feature of a method step. Analogously, aspects
described in the context of a method step also represent a description of a corresponding
block or item or feature of a corresponding apparatus. Some or all of the method steps may
be executed by (or using) a hardware apparatus, like for example, a microprocessor, a
programmable computer or an electronic circuit. In some embodiments, some one or more
of the most important method steps may be executed by such an apparatus.
The inventive encoded watermark signal, or an audio signal into which the watermark
signal is embedded, can be stored on a digital storage medium or can be transmitted on a
transmission medium such as a wireless transmission medium or a wired transmission
medium such as the Internet.
Depending on certain implementation requirements, embodiments of the invention can be
implemented in hardware or in software. The implementation can be performed using a
digital storage medium, for example a floppy disk, a DVD, a Blue-Ray, a CD, a ROM, a
PROM, an EPROM, an EEPROM or a FLASH memory, having electronically readable
control signals stored thereon, which cooperate (or are capable of cooperating) with a
programmable computer system such that the respective method is performed. Therefore,
the digital storage medium may be computer readable.
Some embodiments according to the invention comprise a data carrier having
electronically readable control signals, which are capable of cooperating with a
programmable computer system, such that one of the methods described herein is
performed.

Generally, embodiments of the present invention can be implemented as a computer
program product with a program code, the program code being operative for performing
one of the methods when the computer program product runs on a computer. The program
code may for example be stored on a machine readable carrier.
Other embodiments comprise the computer program for performing one of the methods
described herein, stored on a machine readable carrier.
In other words, an embodiment of the inventive method is, therefore, a computer program
having a program code for performing one of the methods described herein, when the
computer program runs on a computer.
A further embodiment of the inventive methods is, therefore, a data carrier (or a digital
storage medium, or a computer-readable medium) comprising, recorded thereon, the
computer program for performing one of the methods described herein.
A further embodiment of the inventive method is, therefore, a data stream or a sequence of
signals representing the computer program for performing one of the methods described
herein. The data stream or the sequence of signals may for example be configured to be
transferred via a data communication connection, for example via the Internet.
A further embodiment comprises a processing means, for example a computer, or a
programmable logic device, configured to or adapted to perform one of the methods
described herein,
A further embodiment comprises a computer having installed thereon the computer
program for performing one of the methods described herein.
In some embodiments, a programmable logic device (for example a field programmable
gate array) may be used to perform some or all of the functionalities of the methods
described herein. In some embodiments, a field programmable gate array may cooperate
with a microprocessor in order to perform one of the methods described herein. Generally,
the methods are preferably performed by any hardware apparatus.
The above described embodiments are merely illustrative for the principles of the present
invention. It is understood that modifications and variations of the arrangements and the
details described herein will be apparent to others skilled in the art. It is the intent,

therefore, to be limited only by the scope of the impending patent claims and not by the
specific details presented by way of description and explanation of the embodiments
herein.

We Claim:
1. Watermark signal provider (2400) for providing a watermark signal (2440; 101b)
suitable for being hidden in an audio signal (2430; 106) when the watermark signal is
added to the audio signal, such that the watermark signal represents watermark data
(2450; 101a), the watermark signal provider comprising:
a psychoacoustical processor (2410; 102) for determining a masking threshold of the
audio signal; and
a modulator (2420; 307) for generating the watermark signal from a superposition of
sample-shaping functions spaced apart from each other at a sample time interval (Tb)
of a time-discrete representation of the watermark data, each sample-shaping
function being amplitude-weighted with a respective sample of the time-discrete
representation, multiplied by a respective amplitude weight depending on the
masking threshold, the modulator being configured such that
the sample time interval is shorter than a time extension of the sample-shaping
functions; and
the respective amplitude weight also depends on samples of the time-discrete
representation neighboring the respective sample in time.
2. Watermark signal provider according to claim 1, wherein the psychoacoustical
processor is configured to determine the masking threshold independent from the
watermark data and the modulator is configured to generate the watermark signal
iteratively by
preliminarily determining a preliminary amplitude weight based on the masking
threshold independent from the watermark data;
checking as to whether a superposition of the sample-shaping functions using the
preliminary amplitude weight as the respective amplitude weight violates the
masking threshold; and
if the superposition of the sample-shaping functions using the preliminary amplitude
weight as the respective amplitude weight violates the masking threshold, varying
the preliminary amplitude weight so as to obtain a superposition of the sample-

shaping functions using the varied amplitude weight as the respective amplitude
weight.
3. Watermark signal provider according to claim 1 or 2, further comprising a time-
spreader (305) for spreading the watermark data in time in order to obtain the time-
discrete representation.
4. Watermark signal provider according to any of claims 1 to 3, further comprising a
frequency-spreader (303) for spreading the watermark data in a frequency domain in
order to obtain the time-discrete representation.
5. Watermark signal provider according to any of the preceding claims, wherein the
psychoacoustical processor comprises a time/frequency analyzer (501) transferring
the audio signal from the time domain to a frequency domain by means of a lapped
transform using a first window length of approximately the sample time interval.
6. Watermark signal provider according to claim 5, wherein the time/frequency
analyzer is configured to transfer the audio signal from the time domain to the
frequency domain by means of the lapped transform also using a second window
length being shorter than the first window length.
7. Watermark signal provider according to any of the preceding claims, wherein the
time-discreet representation is composed of time-discrete subbands, wherein the
modulator is configured to generate the watermark signal from, for each time-
discrete subband, a superposition of sample-shaping functions spaced apart at the
sample time interval with each sample-shaping function being amplitude-weighted
with a respective sample of the respective time-discrete subband multiplied by a
respective amplitude weight depending on the masking threshold, the sample-
shaping functions of the superposition for a respective time-discrete subband
comprising a carrier frequency at a center frequency of the respective time-discreet
subband.
8. Watermark embedder comprising
a watermark signal provider for providing a watermark signal suitable for being
hidden in an audio signal when the watermark signal is added to the audio signal,
such that the watermark signal represents watermark data, according to any of the
preceding claims, and

an adder for adding the watermark signal and the audio signal to obtain a
watermarked audio signal.
9. Method for providing a watermark signal (101b) suitable for being hidden in an
audio signal (106) when the watermark signal is added to the audio signal, such that
the watermark signal represents watermark data (101a), the method comprising:
determining a masking threshold of the audio signal; and
generating the watermark signal from a superposition of sample-shaping functions
spaced apart from each other at a sample time interval (Tb) of a time-discrete
representation of the watermark data, each sample-shaping function being amplitude-
weighted with a respective sample of the time-discrete representation, multiplied by
a respective amplitude weight depending on the masking threshold, the generation
being performed such that
the sample time interval is shorter than a time extension of the sample-shaping
functions; and
the respective amplitude weight also depends on samples of the time-discrete
representation neighboring the respective sample in time.
10. Watermark embedding method comprising
providing a watermark signal suitable for being hidden in an audio signal when the
watermark signal is added to the audio signal, such that the watermark signal
represents watermark data, according to claim 9, and
adding the watermark signal and the audio signal to obtain a watermarked audio
signal.
11. Computer program having instructions stored thereon for performing, when running
on a computer, a method according to claim 9 or 10.