`
`Qiang Cheng
`
`University of Illinois
`Urbana Champaign,IL
`
`Jeffrey Sorensen
`
`IBM T. J. Watson Research Center
`Yorktown Heights, NY
`sorenj@us.ibm.com
`
`ABSTRACT
`
`The technique of embeddinga digital signal into an audio record-
`ing or image using techniques that render the signal imperceptible
`has received significant attention. Embedding an imperceptible,
`cryptographically secure signal, or watermark, is seen as a poten-
`tial mechanism that may beused to prove ownershipordetect tam-
`pering. While there has been a considerable amountofattention
`devoted to the techniques of spread-spectrum signaling for use in
`image and audio watermarking applications, there has only been a
`limited study for embedding data signals in speech. Speechis an
`uncharacteristically narrow bandsignal given the perceptual capa-
`bilities of the human hearing system. However, using speechanal-
`ysis techniques, one may design an effective data signalthat can be
`usedto hide an arbitrary messagein a speech signal. Also included
`are experiments demonstrating the subliminal channel capacity of
`the speech data embedding technique developedhere.
`
`1. INTRODUCTION
`
`Watermarking is a technique for embedding a cryptographic sig-
`nature into digital content for the purposes of detecting copying or
`alteration of the content. This is accomplished using coding tech-
`niques that hide data within the image or audio content in a manner
`not normally detectable. This paper focuses on an, asyet, largely
`unexplored aspect area of audio watermarks: speech.
`For audio watermarking, Preuss, et. al.
`[8] invent a digital
`information hiding technique for audio using the techniques of
`spread spectrum modulation. Boney, et. al.
`[2] explicitly make
`use of MPEG-1 Psychoacoustic Model to obtain the frequency
`masking values to achieve good imperceptibility. Recently Riuz
`and Deller [9] propose a speech watermarking methodfor the ap-
`plication to the digital speech libraries. These methods have been
`extensively applied for music applications, but embed information
`over a very wide audio band based on humanhearing capabilities.
`A potential attacker need only low-passfilter the resulting signal
`to remove most of the watermarking information.
`Speech differs from music in their acoustic characteristics and
`watermarking requirements. Speech is an acoustically rich signal
`that it uses only a small portion of the human perceptual range.
`Typical speech reproduction hardware, although often the same as
`used with music, includes much lower bit rate channels such as
`telephone or compressed voice “vocoders.” However,
`the same
`analysis techniques employed in such voice coding schemes can
`easily be adapted to create an audio watermarking signal that is
`robustto speech channels. Presented here is a technique for encod-
`ing an additional, arbitrary digital messageinto speechsignals. By
`making use of the well understood techniques of speech analysis,
`
`significantly higher bit rates can be embedded without effecting
`the perceived quality of the recording.
`
`Thedigital hiding technique for speech can be applied to copy-
`right protection for digital speech libraries, audio books,as well as
`covert communication channel. The embedded information may
`be any digital message. Messagesthat can be used to prove author-
`ship require the generation of an appropriate cryptographically se-
`cure digital message and are beyondthe scope ofthis paper. How-
`ever, consult [4] for information on the application of watermarks.
`
`2. VOICEBAND SPREAD SPECTRUM SIGNAL
`
`In contrast to previous work on audio watermarking, the speech
`signal is a considerably narrower bandwidth signal. The long-
`time-averaged powerspectral density of speech indicatesthat the
`signal is confined to a range of approximately 10 Hz to 8 kHz
`[6].
`In order that the watermark survives typical transformation
`of speech signals, including speech codecs,it is importantthat the
`watermark be limited to the perceptually relevant portions of the
`spectra. However,
`the watermark should remain imperceptible.
`Therefore, a spread-spectrum signal with an uncharacteristically
`narrow bandwidth will be used.
`
`Using a direct sequence spread spectrum [3] signal, we wish
`to design a PN sequence with a main side lobe thatfits within a
`typical telephone channel [5], which ranges from 250 Hz to 3800
`kHz.
`In this work, the message sequence and the PN sequence
`are modulated using simple Binary Phase Shift Keying (BPSK).
`The center frequency of the carrier is chosen to be fe = 2025Hz.
`The clock rate of the PN sequence, or chip rate,
`is taken to be
`1775Hz, whichis half of the signal bandwidth. Because the width
`of our watermark is very close to the modulation frequency, it is
`necessary to low passfilter the spread spectrum signal before mod-
`ulation to prevent excessive aliasing. For this, we have chosen to
`use a seventh order Butterworthfilter with a cutoff of 3400 Hz.
`
`illustrates the power spectral density of the water-
`Figure |
`mark signal, with the long-term average speech power spectrum
`(for both a male and female speaker) for illustration. The sim-
`plest implementation of a speech watermark system would involve
`adding this signal, which sounds primarily like radiostatic, to the
`speech signal at the appropriate gain. However, taking advantage
`of our knowledge of the speech signal itself, we are able to em-
`bed a significantly higher gain signal using techniquesthatare the
`subject of the next two sections.
`
`0-7803-7041-4/01/$10.00 ©2001 IEEE
`
`1337
`
`Sony Exhibit 1023
`Sony Exhibit 1023
`Sony v. MZ Audio
`Sony v. MZ Audio
`
`
`
`PowerSpectrumMagnitude(48) 1 &
`xJ:888
`
`
`
`
`
`
`
` 0
`
`
`
`\ SAANfA,ww~YN
`
`os
`
`<
`
`15
`
`2
`
`Frequency
`
`
`
`4.
`
`25
`x 10°
`
`
`
`
`
`~60F
`
`
`
`Watermark Sign
`
`]
`
`\\
`
`
`
`D 8 PowerSpectralDensity(d8)
`
`\a
`
`har
`\
`
`1
`.
`200025
`0030
`Frequency (Hz)
`
`.
`oO
`
`1
`380040
`
`\
`0045
`
`aS pwr 'yl Sey
`se
`
`at ape N
`v
`N
`NasYOY
`\
`
`-J
`
`5000
`
`1
`
`oo
`
`500
`
`:
`100015
`
`1.
`00
`
`& 8
`
`-140]
`
`to)
`
`-180
`
`ae
`
`0
`
`
`
`
`
`
`
`Figure 1: Power spectral densities of the watermark, male voice,
`and female voice.
`
`3. LPC ANAYLSIS AND FILTERING
`
`Ourgoal is to add as much watermark signal energy as possible to
`the speech signal, while still satisfying the constraint that the added
`signal not be perceivable whenlistened to. Most watermarking ap-
`proachesrely on a perceptual model of human hearing. Speechis
`an inherently complex stimuli with rapidly changing spectral char-
`acteristics. Conventional masking effects are most often studied
`for spectral bands outside the range of speech, above 4 kHz. How-
`ever, an effective production model for speech is available. The
`well knowntechnique of linear prediction has proven to be highly
`effective in modeling speech signals.
`In addition, human speech
`perceptionreflects the production system characteristics. Ourfind-
`ings indicate that using the production model can provide excellent
`hiding characteristics.
`In our watermark signal embedding algorithm, the watermark
`signalis filtered to match the overall spectral shape of the speech
`signal. In addition, the linear predictive analysis provides an effec-
`tive dynamic measure of the degree of noise already present in the
`speech signal. Portions of speech that have a highly white spec-
`trum, fricative sounds and the rapidly changing plosives sounds,
`are especially good candidates for embedding additional water-
`mark energy.
`Linear predicative analysis of speech involves computing the
`maximumlikelihood coefficientsofanall-pole filter of the form
`if
`
`A(z) =
`
`Go +a12z~>1+---+ap27P
`
`(1)
`
`Thereis a considerableliterature on the application of linear pre-
`diction to speech signals. For our analysis, we have chosen to use
`the Levinson-Durbin recursive technique for evaluating LPC coef-
`ficients a; from the short-term autocorrelation coefficients.
`The short term autocorrelation can be computed from the win-
`dowed speech frame s(t) as
`N-1
`— S> s(n)s(n — #)
`n=1
`
`which,
`
`TO
`ry
`
`Pi
`To
`
`tae
`ts
`
`[Tp-1
`Tp-2
`
`Tp-1
`
` Tp-2 + To
`
`a4
`a2
`
`Ap
`
`T1
`r2
`
`Tp
`
`Figure 2: Power spectrum of a segment of speech and spectrum of
`LPC-shaped watermark signal.
`
`which, in vector notation can be represented by
`
`Ra=r
`
`The prediction residual energy, or the average squared-error can
`be computed as
`
`E=a'Ra
`
`is a measure of the “predictability” of the speech signal, and an
`effective measure of the noise content.
`Beforefiltering the watermark signal using theall-pole filter,
`a bandwidth expansion operation is performed. This movesall of
`the poles closer to the center ofthe unit circle, increasing the band-
`width of their respective resonances. The vocaltract filter often
`tends to have quite narrow spectral peaks. Due to masking phe-
`nomena, sounds near these peaks are unlikely to be perceived by
`the listener. Therefore, by increasing the bandwidth of formantre-
`sponses, larger overall watermark signal gains should betolerable.
`The bandwidth parameter ¥ is used to adjust the LPC coefficients
`— a
`a; = aiy
`
`where y may be chosen between 0 and1.
`Figure 2 shows the power spectrum of a segment of speech,
`and the spectrum of the watermark signal that results afterfiltering
`using the spectral envelope of the speech segment.
`
`4. WATERMARK SIGNAL GAIN
`
`The instantaneous watermark gain is dynamically determined to
`match the characteristics of the speech signal. In the simplest case,
`whenlittle speech energy is present(i.e. during silence) the wa-
`termark is added using a fixed gain threshold. This is selected so
`that the watermark becomesthe effective noisefloorof the record-
`ing. Perceptually, a small amountof noise is always expected in a
`recording and the watermarksignalis not atypical of such record-
`ing noise. In many applications, silence may not be transmitted or
`might be by coded using extreme compression.
`In these circum-
`stances, designers should choose anerror correcting code (such as
`a convolutional code) with the proper characteristics so that the
`message may be recovered despite these losses.
`The normalized per sample speech energy EFfor one frameis
`N
`¢
`Es = % oy, 8 (n) = Fro.
`
`1338
`
`
`
`
`
`
`
`
`
`BitErrorProbability
`
`
`
`BX8&BitErrorProbability
`
`0
`
`20
`
`60
`40
`Frame Rate(ms)
`
`0
`
`100
`
`500
`
`1500
`4000
`Message Bit Rate (bits/sec)
`
`2000
`
`(%)
`
` 200
`
`Figure 4: Bit Error Probability versus Frame Rate, and Bit Error
`Probability versus Message Bit Rate.
` 250
`—*- Female Speaker
`
`
`
`
`
`
`
`WatermarkingChannelCapacity(bits/sec)
`
`-©- Male Speaker
`
`50
`
`0
`
`500
`
`1500
`1000
`MessageBit Rate(bits/sec)
`
`2000
`
`
`
` 1 L 1 1 L 1 1 1
`
`
`
`
`
`
`
`0
`04
`02
`03
`04
`os
`06
`O7
`08
`09
`1
`
`
`Energy
`
`x10”
`ey ‘
`:
`T
`1
`t
`Constant, Predictor Error and EnergyTerms
`Constant + Predictor Eyror
`\
`\
`41Cons tant Tern
`Xe
`
`3
`\ \
`
`
`
`\
`
`nn
`YNoap
`sae
`Withee
`
`02
`03
`04
`05
`0
`4
`0.
`Time(s)
`
`.
`
`4
`
`|
`
`06
`
`07
`
`08
`
`09
`
`1
`
`
`
`WatermarkAdditive
`
`5
`
`1
`
`Figure 3: A segmentof speech and the corresponding watermark
`gains.
`
`The watermark gain in each frame can be determined bythe
`linear combinationofthe gains for silence, normalized per sample
`residual energy FE, and normalized per sample speech energy Es,
`
`g(t) =Ao+ AEF AEs,
`
`(2)
`
`which is designed to maximize the strength of the watermark sig-
`nals without incurring perceptual degradations. Figure 3 shows a
`segment of speech and the embedded watermarksignal. The re-
`sulting watermarked speech is shownalso in Figure 3. Listening
`test demonstrates that the watermarked speechis indistinguishable
`from the original speech with this watermark gain. If the gain is
`increased further, there will be “hoarseness” in the watermarked
`speech. Thoughit hardly affects the naturalness of the voice, the
`difference with the original speech is indeed perceptible.
`
`5. WATERMARK DETECTION
`
`Atthe receiving end, the received signal ro(t) is given by
`N
`ro(t) = Y> w(t) + s(t) + Lo(t),
`
`t=1
`
`63)
`
`where w(t) is the LPC-shaped watermark signal, s(£) is the orig-
`inal speech signal, and Io(t) is somedeliberated attacksordigital
`signal processing. Weestimate the LPC coefficients from the re-
`ceived signal, and then take the inverse LPCfiltering of ro(t) to
`get r(t). After inverse LPCfiltering, voiced speech becomesperi-
`odie pulses, and unvoiced speech becomes whitened noise. Asis
`typical for speech processing, we modelthe inversefiltered s(t) as
`White Gaussian Noise (WGN). Inverse LPCfiltering decorrelates
`the speech samples s(t) as well as equalizes the watermarksignal
`w(t). A correlation receiver,
`
`N
`Hy
`Y= d(t)r(t) > 0,
`t=1
`
`(4)
`
`Figure 5: Watermarking Channel Capacity versus Message Bit
`Rate
`
`desired robustnessproperty. The decodingrule is a maximum like-
`lihood decisionrule, which is also a minimum probability-of-error
`rule since 0 and 1 in the message are sent with equal probabilities.
`The problem of synchronization when the original message is
`not available is beyond the scope of this paper. However, the PN
`sequence used in the spread spectrum modulation can be used to
`drive a phase locked loop during decoding. The techniques pre-
`sented in [3] [8] can be used in our framework for synchronization
`purposes.
`:
`
`6. EMBEDDED CHANNEL CAPACITY
`
`A set of simulation experiments were performed to demonstrate
`the relationship betweenthe frame size and messagerate (1 bit per
`frame) and thebit error probability, as shown in Figure 4.
`The spread spectrum signal, when addedto the original speech,
`can be considered as a noisy communication channel, called the
`watermarking channel. The watermarkis the content of the trans-
`mitted message. Without loss of generality, the message is con-
`sidered to be a binary signal with equal probability for 0 and 1.
`The watermark channel is binary symmetric The channel capac-
`ity, which is the theoretical maximumrate for data transmission,is
`defined for the watermarking channel [1]:
`
`gives us optimumdetection performance in AWGN [7], where NV
`is the length of a frame, in which one messagebit is embedded,
`d(t) is the despreading function, which is the synchronized, BPSK
`modulated spreading function for the current frame. The correla-
`tion with d(t) can average outthe interference, thus providing the
`
`C = R(1 + plogop + (1 — p)loge(1 — p)),
`
`(5)
`
`where p is the crossover probability, R is the message bit rate.
`The simulation results for the watermarking channel capacity are
`plotted in Figure 5. For a binary symmetric channel, the chan-
`
`1339
`
`
`
`tioning system can be built using the data embedding algorithm
`presented here, where the text transcription of the speech would
`be hidden in the speechitself. In addition, in-band signaling ap-
`plications, typically done using dual tone *touch-tone”signals can
`be replaced with embedded contro] signals, suggesting novelsi-
`multaneous voice and data applications. For the purposeofside-
`information embedding,
`thereis little threat from intentional at-
`tacks. Thus, a larger capacity of information can be communicated
`with less dependency on the redundancy oferror correct codings.
`
`9. REFERENCES
`
`[1] R. E. Blahut. Principles and Pratice of Information Theory.
`Addison-Wesley Publishing Company, 1987.
`
`[2]
`
`L. Boeny, A. H. Tewfik, and K. N. Hamdy. Digital watermarks
`for audio signals.
`In Proc. of Multimedia 1996, Hiroshima,
`1996.
`
`G. R. Cooper and C. D. McGillem. Modern Communica-
`tions and Spred Spectrum. McGraw-Hill Book Company, New
`York, 1986.
`
`F. Hartung and M. Kutter. Multimedia watermarking tech-
`niques. In Proceedings of the IEEE, vol. 87, July, 1999.
`
`C. Jankowski, A. Kalyanswamy, S. Basson, and J. Spitz.
`Ntimit: A phonetically balanced, continuous speech,
`tele-
`phone bandwidth speech database.
`In JCASSP, pages 109-
`112, Albuquerque, NM, 1990.
`N. S. Jayant and P. Noll. Digital Coding of Waveforms. Pren-
`tice Hall, Inc., Englewood Cliffs, New Jersey, 1984.
`
`H. V. Poor. An Introduction to Signal Detection and Estima-
`tion. Springer-Verlag, New York, 1994.
`
`R. Preuss, S. Roukos, A. Huggins, H. Gish, M. Bergamo, and
`P. Peterson. Embedding Signalling. US Patent 5319735, 1994.
`
`R. J. Ruiz and J. R. Deller. Digital watermarking of speech
`signals for the national gallery of the spoken word. In JCASSP,
`Turkey, 2000.
`
`1340
`
`
`
`
`Speec
`Watermark
`Bit
`Bit
`Rate
`Reliability
`706 kbps
`74.05%
`128 kpbs
`71.58%
`32 kbps
`68.65%
`13 kpbs
`61.23%
`
`
`
` ;
`
`Compression
`Scheme
`16 bit linear PCM
`16 bit linear PCM
`IMA ADPCM
`GSM 6.10
`
`Speech
`Bandwidth
`22 kHz
`4 kHz
`4 kHz
`4kHz
`
`Table 1: Watermarking Attacks by Voice Compression
`
`nel capacity is achievable [1]. That is, transmission codes can be
`designed for reliable communication underoratthisrate.
`The plot shows that the frame size needs to be small when
`high channel capacity is desired. However, the LPC prediction
`suffers when the frame size is too small, which makes LPC shap-
`ing less effective. And also the degradation of the watermarking
`channel due to attacks is more severe for smaller frame, see Sec-
`tion 7. Therefore, there is an intrinsic tradeoff between channel
`capacity and survivability of watermark. To achieve high channel
`capacity, good LPC predictability, and reasonable survivability si-
`multaneously we have chosen 800 bits per second as our message
`embeddingrate.
`
`7. ROBUSTNESS
`
`Watermarked media is subject to a variety of attacks. With images,
`images may be cropped,rotated, filtered, or otherwise changed.
`Audio signals are less subject to these types of manipulations, as
`the human perceptual system is quite sensitive to changes in au-
`dio signals. However, speech signals may be affected by transfor-
`mations that include: analog to digital and digital to analog con-
`versions, filtering, re-equalization, changes in playback rate, and
`compression. The algorithm presented here puts all of the water-
`marksignal in the most perceptually important areas of the speech
`signal. Therefore, primitive attempts to remove the watermark by
`filtering are almost certain to proveineffective.
`In order to demonstrate the robustness of the data embedding
`scheme, wehave used an analog reproduction system to simulate a
`crude attemptat duplication. A recording is made at 8 kHz, signif-
`icantly reducing the bandwidth, and then the signal is re-sampled
`at the original rate. This could be considered similar to recording
`across a telephone channel, although no explicit telephone net-
`work equalization was applied. Finally, these 8 kHz recording
`were compressed and decompressedusingthe typical speech com-
`pression algorithms IMA ADPCMand GSM 6.10. Theresults are
`summarized in Table 7.
`
`8. APPLICATIONS AND FUTURE WORK
`
`This paper presents a technique for embedding an arbitrary mes-
`sage In a speech signal.
`In order to provide a complete water-
`marking application, one must choose a messagethat provides the
`appropriatecryptographic properties, such as proof of authenticity
`or ownership. In this respect, the embedding algorithm presented
`here can be used with nearly any comparable application. For ex-
`ample,it can be applied to the copyright of the language-learning
`CD's, audio books, recorded teleconferencing data, digital speech
`libraries [9] and Internet radio broadcasts,etc.
`In addition, a speech data embeddingalgorithm suggests some
`new andpossibly unique applications. For example, a closed cap-
`
`