`
`ir V
`ouoG
`
`1i
`oaaMMM
`
`WOO
`
`IPR No. 2017-00626
`Apple Inc. v. Andrea Electronics Inc. - Ex. 1029, p. Exhibit Cover
`
`
`
`UNITED STATES INTERNATIONAL TRADE COMMISSION
`WASHINGTON, D.C.
`
`Before The Honorable Thomas B. Pender
`Administrative Law Judge
`
`In the Matter of
`
`CERTAIN AUDIO PROCESSING HARDWARE
`AND SOFTWARE AND PRODUCTS
`CONTAINING THE SAME
`
`Investigation No. 337-TA-949
`
`DECLARATION OF SCOTT C. DOUGLAS, Ph.D.
`IN SUPPORT OF COMPLAINANT ANDREA ELECTRONIC CORPORATION'S
`INITIAL CLAIM CONSTRUCTION BRIEF
`
`IPR No. 2017-00626
`Apple Inc. v. Andrea Electronics Inc. - Ex. 1029, p. Cover
`
`
`
`
`
`I, Scott C. Douglas, Ph.D., declare as follows:
`
`1.
`
`I have been retained by Complainant Andrea Electronics Corporation (hereinafter
`
`"Andrea" or "Complainant") to provide expert opinion and testimony in connection with the
`
`above captioned Investigation. In particular, I have been asked by Andrea to provide expert
`
`opinions with regards to the construction of certain claim terms of Andrea's U.S. Patent Nos.
`
`6,363,345 ("the '345 Patent"), 6,049,607 ("the '607 Patent"), and 6,377,637 ("the '637 Patent").
`
`The expert opinions that I set forth in my declaration are based upon my knowledge in the field,
`
`the patents at issue in this Investigation, the file histories of the patents at issue in this
`
`investigation, and the various texts that I rely upon in my declaration. I am being compensated
`
`at a rate of $550 per hour. My compensation is in no way dependent upon or contingent upon
`
`the opinions and testimony that I render during the course of this Investigation.
`
`2.
`
`1 am currently a professor in the Department of Electrical Engineering at the
`
`Bobby B. Lyle School of Engineering at Southern Methodist University. I have been a professor
`
`in the Department of Electrical Engineering at Southern Methodist University since August
`
`1998. I have taught, and continue to teach, courses to undergraduate and graduate level students
`
`in the areas of signal processing, including adaptive filtering and adaptive arrays. My research at
`
`Southern Methodist University is focused in the areas of acoustic signal processing, active noise
`
`control, adaptive filtering, array processing, multichannel blind deconvolution and source
`
`separation.
`
`3.
`
`Prior to my position at Southern Methodist University, I was an assistant
`
`professor in the Department of Electrical Engineering at the University of Utah. I taught courses
`
`to undergraduate and graduate level students in the areas of signal processing, including digital
`
`signal processing, adaptive filtering, and active noise control. In addition to teaching, I also
`
`performed research in the areas of adaptive filtering, active noise control, multichannel blind
`
`-1-
`
`IPR No. 2017-00626
`Apple Inc. v. Andrea Electronics Inc. - Ex. 1029, p. 1
`
`
`
`
`
`deconvolution and source separation, and hardware implementations of adaptive signal
`
`processing systems.
`
`4.
`
`I have been a member of the Institute of Electrical and Electronics Engineers
`
`since 1988, and am currently a Senior Member. I have been an Associate Editor of the IEEE
`
`Transactions on Signal Processing and IEEE Signal Processing Letters. I have had leadership
`
`roles in IEEE organizational activities, including conference and workshop organization, and I
`
`have served on three Technical Committees of the IEEE Signal Processing Society and held
`
`leadership positions of Secretary or Chair of some of these committees. In 2010,1 was the
`
`General Chair and the organizer of the IEEE International Conference on Acoustics, Speech, and
`
`Signal Processing, the premier yearly IEEE conference series on all aspects of signal processing
`
`theory, methods, and applications, and I have published in and attended this conference every
`
`year it has been offered since 1990. 1 was the recipient of the Best Paper Award in Audio and
`
`Electroacoustics of the IEEE Signal Processing Society in 2003.
`
`5.
`
`1 have written several book chapters related to adaptive filters, microphone arrays,
`
`blind deconvolution, and source separation. I was section editor of the Adaptive Filters portion
`
`of The Digital Signal Processing Handbook, Vijay Madisetti and Douglas Williams, eds. (Boca
`
`Raton, FL: CRC/IEEE Press, 1998), and authored one chapter and co-authored another chapter
`
`on adaptive filters for this text. I co-authored, with Shun-ichi Amari, the book chapter entitled
`
`"Natural Gradient Adaptation," in Unsupervised Adaptive Filtering Vol. I: Blind Signal
`
`Separation, Simon Haykin, ed., (New York: Wiley, 2000), and I co-authored, with Simon
`
`Haykin, the book chapter entitled "Relationships Between Blind Deconvolution and Blind
`
`Source Separation," in Unsupervised Adaptive Filtering, Vol. IT Blind Deconvolution, Simon
`
`Haykin, ed., (New York: Wiley, 2000). I wrote the book chapter entitled, "Blind Separation of
`
`-2-
`
`IPR No. 2017-00626
`Apple Inc. v. Andrea Electronics Inc. - Ex. 1029, p. 2
`
`
`
`
`
`Acoustic Signals,' appearing in Microphone Arrays: Techniques and Applications, Michael
`
`Brandstein and Darren Ward, eds., (New York: Springer-Verlag, 2001). I co-authored, with
`
`Malay Gupta, the book chapter entitled, "Convolutive Blind Source Separation for Audio
`
`Signals," in Blind Speech Separation, Shoji Makino, Te-Won Lee, and Hiroshi Sawada, eds.
`
`(New York: Springer, 2007).
`
`6.
`
`1 received my bachelors, masters, and doctorate degrees in electrical engineering
`
`from Stanford University. For my doctorate degree, the focus of my studies were in the area of
`
`signal processing, adaptive filters, and statistical estimation and detecting. I received my
`
`doctorate degree in 1992. A copy of my curriculum vitae is attached as Exhibit 1.
`
`I.
`
`LEGAL STANDARDS
`
`7.
`
`I am not an attorney, but I have been informed of the following standards
`
`regarding claim construction:
`
`•
`
`Claim construction begins with the words of the claim itself, which generally receive
`
`their ordinary and customary meaning as understood by a person of ordinary skill in
`
`the art at the time of the invention in the context of the specification and prosecution
`
`history. Phillips v. A WHCorp., 415 F.3d 1303, 1312-13 (Fed. Cir. 2005) (en banc).
`
`To ascertain the ordinary and customary meaning of the claims, courts consider the
`
`intrinsic record, including the claims, the specification, and the prosecution history.
`
`Id. at 1314. Claim terms "can be defined only in a way that comports with the
`
`instrument as a whole[]" and must be read "in the context of the entire patent[.]"
`
`Markman v. Westview Instruments, Inc., 517 U.S. 370, 389 (1996). It is the claims
`
`that delimit a patentee's right to exclude, and therefore it is not proper to import
`
`limitations from the specification into the claims. Varco, L.P. v. Pason Sys. USA
`
`Corp., 436 F.3d 1368, 1373 (Fed. Cir. 2006). A patentee need not describe in the
`
`-3-
`
`IPR No. 2017-00626
`Apple Inc. v. Andrea Electronics Inc. - Ex. 1029, p. 3
`
`
`
`
`
`specification every conceivable and possible future embodiment of his invention."
`
`CCS Fitness, Inc. v. Brunswick Corp., 288 F.3d 1359, 1366 (Fed. Cir. 2002) (internal
`
`quotation marks and citation omitted). On the other hand, "a claim interpretation that
`
`excludes a preferred embodiment from the scope of the claim is rarely, if ever,
`
`correct." On-Line Techs., Inc. v. Bodenseewerk Perkin-Elmer GmbH, 386 F.3d 1133,
`
`1138 (Fed. Cir. 2004).
`
`In addition to the specification and claims, the court may also consider the
`
`prosecution history, which, "[l]ike the specification, .
`
`. provides evidence of how the
`
`.
`
`PTO and the inventor understood the patent." Phillips, 415 F.3d at 1312 (citation
`
`omitted). In addition, "[a] court can look to the prosecution history of related patents
`
`for guidance in claim construction[.]" Aventis Pharms. Inc. v. Amino Chems. Ltd.,
`
`715 F.3d 1363, 1375 (Fed. Cir. 2013) (citation omitted).
`
`• Courts may also consider extrinsic evidence, e.g., inventor testimony, dictionaries,
`
`and treatises, when intrinsic record alone is insufficient to support proper
`
`constructions. Phillips, at 1317-18. Expert testimony is often helpful to illuminate
`
`complex technical issues and provide a foundation for the viewpoint of one of
`
`ordinary skill in the relevant art. Id. at 1318 ("We have also held that extrinsic
`
`evidence in the form of expert testimony can be useful to a court for a variety of
`
`purposes, such as to provide background on the technology at issue, to explain how
`
`an invention works, to ensure that the court's understanding of the technical aspects
`
`of the patent is consistent with that of a person of skill in the art, or to establish that a
`
`particular term in the patent or the prior art has a particular meaning in the pertinent
`
`field."). As the Federal Circuit explained, "[t]he construction that stays true to the
`
`-4-
`
`IPR No. 2017-00626
`Apple Inc. v. Andrea Electronics Inc. - Ex. 1029, p. 4
`
`
`
`
`
`claim language and most naturally aligns with the patent's description of the
`
`invention will be, in the end, the correct construction." Id. at 1316 (internal quotation
`
`marks and citation omitted).
`
`8.
`
`Likewise, I have been informed of the following standards regarding
`
`indefiniteness:
`
`A patent must "conclude with one or more claims particularly pointing out and
`
`distinctly claiming the subject matter which the applicant regards as [the]
`
`invention." 35 U.S.C. § 112, ¶ 2 (2006). A claim fails to satisfy this statutory
`
`requirement and is thus invalid for indefiniteness only if its language, when read
`
`in light of the specification and the prosecution history, "fail[s] to inform, with
`
`reasonable certainty, those skilled in the art about the scope of the invention."
`
`Nautilus, Inc. v. Biosig Instruments, Inc., 134 S. Ct. 2120, 2124 (2014). This
`
`standard allows for some amount of uncertainty, as absolute precision in claim
`
`drafting is unattainable. Id at 2128-29. Instead, indefiniteness problems arise
`
`where the claim language "might mean several different things and 'no informed
`
`and confident choice is available among the contending definitions'." Interval
`
`Licensing LLC v. AOL, Inc., 766 F.3d 1364, 1371 (2014) (citing Nautilus, 134 S.
`
`Ct. at 2130 & n. 8 (quoting Every Penny Counts, Inc. v. Wells Fargo Bank, NA.,
`
`2014 U.S. Dist. LEXIS 28106, 2014 WL 869092, at *4 (M.D. Fla. Mar. 5,
`
`2014))). Claim drafting flaws, such as lack of antecedent basis, do not
`
`automatically render claims indefinite. See, e.g., Trover Grp., Inc. v. Dedicated
`
`Micros USA, 2015 U.S. Dist. LEXIS 33876 at *28 (E.D. Tex. March 19, 2015)
`
`-5-
`
`IPR No. 2017-00626
`Apple Inc. v. Andrea Electronics Inc. - Ex. 1029, p. 5
`
`
`
`
`
`(citing Nautilus, 134 S. Ct. at 2124). A claim may be reasonably clear to a person
`
`of skill in the art even in the presence of drafting flaws. Id.
`
`H.
`
`BACKGROUND OF THE TECHNOLOGY
`
`9.
`
`The technology at issue in this Investigation generally relates to the processing of
`
`audio signals, and more specifically to canceling noise in audio signals.
`
`A.
`
`10.
`
`Audio Signals
`
`Audio signals are a representation of sound. Sound is a vibration that propagates
`
`as a pressure wave through a medium (e.g., air). These physical sound waves can be represented
`
`in terms of electrical voltage, for example, when picked up by a microphone. A microphone
`
`typically includes a membrane which vibrates when the sound wave (vibrations) impact the
`
`membrane. The vibrations of the membrane are converted into electrical energy and measured in
`
`terms of electrical voltage. When a loud sound (which carries more energy) hits the microphone
`
`membrane, it results in more vibration in the membrane, which translates into a higher measure
`
`of electrical voltage. Likewise, when a soft sound (which carries less energy) hits a microphone
`
`membrane, it results less vibration in the membrane, which translates into a lower measure of
`
`electrical voltage. These measures of electrical voltage, collected over time, correspond to the
`
`audio signal, and can be plotted to show the waveform of the audio signal. Below is an example
`
`of the waveform of an audio signal:
`
`-6-
`
`IPR No. 2017-00626
`Apple Inc. v. Andrea Electronics Inc. - Ex. 1029, p. 6
`
`
`
`
`
`iie
`
`Figure 1: Illustration of an audio signal. The audio signal corresponds a signal generated from
`a tone.
`
`The oscillations in the waveform correspond to the physical vibration of the membrane in a
`
`microphone.
`
`11.
`
`Typically, the audio signals generated from a microphone are analog signals.
`
`Analog signals are continuous signals that vary over time. These analog signals can be analyzed
`
`and modified using analog systems (e.g., usually consisting of analog amplifiers, resistors,
`
`capacitors). Analog systems, however, are relatively limited in function and often require
`
`substantial hardware redesign to incorporate additional functionality. Digital systems allow for
`
`much greater functionality, as they are typically implemented in general purpose computers that
`
`can be programmed with software to achieve the desired functionality. As such, audio signals
`
`are typically processed using digital systems.
`
`12.
`
`In order to process audio signals in digital systems, they must first be converted
`
`into digital signals. The conversion of an analog signal into a digital signal is accomplished by
`
`-7-
`
`IPR No. 2017-00626
`Apple Inc. v. Andrea Electronics Inc. - Ex. 1029, p. 7
`
`
`
`
`
`use of an analog-to-digital converter (A/D converter). The A/D converter samples the input
`
`signal by taking measurements of the amplitude of the analog signal at regular time intervals.
`
`This sampling is illustrated, for example, in the diagram below:
`
`ne
`
`Figure 2: Illustration of the sampling of an analog signal.
`
`Unlike analog signals, digital signals allow for the representation of a signal using a discrete and
`
`finite number of points. Because of the discrete nature of digital signals, the digitized samples
`
`may not exactly match the values of the analog signals, as shown for example in Figure 2, above,
`
`which shows that the digital samples do not line up exactly with the analog counterpart. But it is
`
`the discrete nature of digital signals that allows each point of the digital signal to be stored into
`
`discrete memory locations in the digital system.
`
`13.
`
`Higher sampling rates can be used by the A/D converter to provide for a more
`
`accurate representation of the analog signal in digital form. However, using a higher sampling
`
`-8-
`
`IPR No. 2017-00626
`Apple Inc. v. Andrea Electronics Inc. - Ex. 1029, p. 8
`
`
`
`
`
`rate requires increased memory (to store the additional measurements of the amplitude of the
`
`analog signal) and processing power (to process the additional data).
`
`14.
`
`Lower sampling rates can also be used, but too low of a sampling rate may result
`
`in aliasing. When aliasing occurs, the sampled signal becomes indistinguishable from other
`
`signals and can cause unwanted distortions or artifacts in the signal when it is converted back
`
`into analog form. An example of an aliased signal is shown below:
`
`ne
`
`Figure 3: Example of an aliased signal. The blue signal corresponds to the original signal,
`while the red signal shows the aliased signal that resulted from too low of a sampling rate.
`
`To avoid aliasing, the signal should be sampled at a sampling rate that is at least two times the
`
`highest frequency that appears in the signal. For example, if the highest frequency that appears
`
`in the signal is 10kHz, the signal should be sampled at a rate of at least 20kHz to avoid aliasing.
`
`This rule is referred to in the art as the Nyquist theorem or the sampling theorem.
`
`-9-
`
`IPR No. 2017-00626
`Apple Inc. v. Andrea Electronics Inc. - Ex. 1029, p. 9
`
`
`
`
`
`B.
`
`Analyzing Digital Audio Signals
`
`15.
`
`Once the signal is digitized, it can be processed by a digital system. Examples of
`
`processing a digital audio signal include passing the digital signal through a low pass filter to
`
`filter out unwanted high frequency components (e.g., noise) or adjusting the gain (loudness) of
`
`certain portions of the input signal. These relatively simple audio processing operations,
`
`however, are sometimes insufficient to achieve the desired processing.
`
`16.
`
`One convenient class of representations of a digital signal uses the concepts of
`
`frequency analysis that are analogous to the biological systems that are used in our perception of
`
`sound. Our ears are able to pick out individual tones, or frequencies, of a sound due to the
`
`physical construction of the human hearing system and the way sound propagates as it impinges
`
`on our ears.
`
`17.
`
`Engineers have developed ways to process audio signals that can extract
`
`individual frequency elements of a sound in a similar fashion, but with the advantage that these
`
`frequency elements can be precisely manipulated and recombined via high-speed computing to
`
`make new analog signals that can be very different from their original physical counterparts. In
`
`this way, a much more sophisticated processing strategy can be achieved.
`
`18.
`
`Extracting a frequency element involves using a signal processing system called a
`
`filter bank. A filter bank is a set, or bank, of frequency-selective filters that are each tuned to a
`
`specific set of frequencies where a portion of an audio signals energy might be. A simple
`
`example of a signal filtered through a filter bank is illustrated below:
`
`-10-
`
`IPR No. 2017-00626
`Apple Inc. v. Andrea Electronics Inc. - Ex. 1029, p. 10
`
`
`
`/
`
`Filter 1
`
`Filter D2--~, _;
`
`A
`V
`
`A
`
`Filter 3
`
`Figure 4: Illustration of a filter bank filtering a signal into its constituent components.
`
`In particular, Figure 4 illustrates a filter bank consisting of three filters (Filters 1 - Filter 3).
`
`Filter I filters out the highest frequency components of the signal; Filter 2 filters the medium
`
`frequency components of the signal; and Filter 3 filters the lowest frequency components of the
`
`signal. By filtering the input signal into its constituent components, it becomes easier to process
`
`individual components of the signal (e.g., unwanted high frequency components of a signal).
`
`19.
`
`Depending on the processing goals, the chosen set of frequencies filtered out by
`
`each filter in the filter bank can be different — they can overlap, for example, or they can span
`
`different amounts of the audio signal's frequency range - but collectively, they typically cover
`
`IPR No. 2017-00626
`Apple Inc. v. Andrea Electronics Inc. - Ex. 1029, p. 11
`
`
`
`
`
`the entire frequency range, or bandwidth, of the signal being processed. The filter bank produces
`
`frequency elements that are frequency-indexed time-domain samples at the output of each filter
`
`of the filter bank, where the sampling rate of each of the frequency elements can be different,
`
`and is usually lower, than that of the original time-domain signal. A well-known filter bank uses
`
`the Discrete Fourier Transform (DFT) as a basis for its filters' coefficients. Such a DFT filter
`
`bank may be realized using the canonical filter bank structure - a set of parallel, frequency-
`
`selective filters that operate independently. It may sometimes be advantageous to share results
`
`of arithmetic operations across filters in a bank. DFT filter banks may do so via a set of
`
`algorithms collectively known as the Fast Fourier Transforms (FFTs.)
`
`20.
`
`These frequency elements are then collectively used in a specific processing
`
`strategy to achieve a goal, such as signal enhancement or noise removal. After processing, these
`
`elements are then combined by a second processing system that reconstructs a single signal from
`
`the frequency-domain elements. The processing in this reconstruction step looks somewhat like
`
`the reverse of the filter bank, but its mathematical form is usually specifically-constructed to
`
`maintain an accurate representation of the original digital signal if the frequency components are
`
`left unchanged. The reconstruction system processes collections of samples periodically in time
`
`but produces one time-domain output signal from these samples.
`
`21. When processing the frequency elements of a signal, it is often important to
`
`measure the level of the signal in a particular frequency range and at a particular point in time.
`
`The level of the signal measures the average height of a signal, and it is either positive or zero.
`
`A zero-level signal is only possible for a signal that is zero at all times. Calculation typically
`
`involves time-averaging of the signal samples, but we cannot simply average the samples
`
`together to get the signal level. Doing so using the samples from the signal in Figure 5 below
`
`-12-
`
`IPR No. 2017-00626
`Apple Inc. v. Andrea Electronics Inc. - Ex. 1029, p. 12
`
`
`
`
`
`would give an answer that is near zero, even though the signal is not zero everywhere (and in
`
`fact could be negatively-valued at some points in time).
`
`iJ
`
`Time
`
`Figure 5: Example of a period sine wave.
`
`Instead, we need some other way to assess the size of these samples, which due to the symmetry
`
`of the signal about the time axis is the same as looking at the deviation of the samples away from
`
`zero. The magnitude of a signal is a measure of this signal deviation away from zero. The
`
`primary operation in assessing signal magnitude is a calculation of the sample distance away
`
`from zero, which for real-valued signals is done by a process called rectification which reflects
`
`the negative parts of the signal over the time-axis to the positive side, as shown in Figure 6,
`
`below.
`
`-13-
`
`IPR No. 2017-00626
`Apple Inc. v. Andrea Electronics Inc. - Ex. 1029, p. 13
`
`
`
`
`
`0
`
`5
`Time
`
`10
`
`Figure 6: Magnitude of the sine wave in Figure 5 calculated by rectification.
`
`Some signal processing structures, such as certain types of filter banks like the DFT filter bank,
`
`produce complex-valued signals at their outputs. Complex-valued signals have both a real
`
`component and an imaginary component for each signal sample, and is one way of representing
`
`the information contained in the signal. For complex-valued signals, the magnitude is calculated
`
`using both the real and the imaginary parts of each signal sample. For either real-valued or
`
`complex-valued signals, it is a measure of the signal deviation away from zero that is important
`
`to determining the signal level as opposed to the scale of the deviation, and many possible
`
`strategies for computing this deviation can be employed, such as using the squares of the signal
`
`samples as opposed to their absolute values. The overall scaling of signal levels are often
`
`application-dependent, and for audio signals, it is common to employ a logarithmic scale such as
`
`the decibel [dB] scale to better match the perceptual characteristics of loudness in human
`
`hearing. The decibel scale changes the way the signal level is represented, but it keeps the
`
`relative relationships between large and small signal levels intact.
`
`-14-
`
`IPR No. 2017-00626
`Apple Inc. v. Andrea Electronics Inc. - Ex. 1029, p. 14
`
`
`
`
`
`C.
`
`22.
`
`Adaptive Filtering
`
`An adaptive filter is a digital signal processing system that, in its simplest form,
`
`models the relationship between an input signal and a desired response or main signal. Adaptive
`
`filters are used for many different applications in signal processing, including noise removal,
`
`signal enhancement, beamforming, and echo cancellation. The key feature of an adaptive filter
`
`is its ability to tune or adapt its capabilities according to the input and main signals fed to the
`
`filter.
`
`23.
`
`An adaptive filter has a set of adjustable filter coefficients, or weights, that
`
`describe the current relationship between the main and input signals. These filter coefficients are
`
`used to produce an output sample for each sample time of the input signal according to a chosen
`
`modeling structure. For example, if a finite-impulse-response (FIR) filter model is used, then the
`
`output signal is a weighted sum of the input signal samples over a particular time interval, where
`
`the weighting values are the filter coefficients themselves. Other filtering models are also
`
`possible and offer different modeling capabilities.
`
`24.
`
`These filter coefficients are adjusted using an error signal that is computed by
`
`subtracting the output of the adaptive filter from the main signal. The general goal of the
`
`adjustment procedure is to model the portion of the input signal in the main signal as best as
`
`possible, and this modeling is achieved by attempting to reduce the level of the error signal at
`
`each sample time by adjusting the filter coefficients. If the main signal consists largely of
`
`components that are closely-related to the input signal, then the adjustment procedure causes the
`
`level of the error signal to generally decrease over time.
`
`25.
`
`Many procedures or algorithms are available for adjusting the filter coefficients of
`
`the adaptive filter. By far the most popular algorithm is the Least-Mean-Square (LMS)
`
`algorithm, originally developed by Widrow and Hoff in 1959 and made popular for adaptive
`
`-15-
`
`IPR No. 2017-00626
`Apple Inc. v. Andrea Electronics Inc. - Ex. 1029, p. 15
`
`
`
`
`
`noise cancellation in a famous IEEE paper by Widrow and his colleagues in 1975. For a chosen
`
`filter model, the LMS algorithm has a single adjustable parameter, called the step size, that
`
`controls the rate of adaptation and the performance of the filter. Variants of this algorithm
`
`include the Normalized Least-Mean-Square (NLMS) algorithm, which uses a normalized step
`
`size parameter that makes the algorithm less sensitive to the level of the input signal.
`
`III. OVERVIEW OF THE ASSERTED PATENTS
`
`A.
`
`26.
`
`U.S. Patent No. 6,363,345
`
`Andrea's '345 Patent is directed towards a system, method, and apparatus for
`
`cancelling or reducing noise using spectral subtraction. ('345 Patent, col. 1:19-21.) Spectral
`
`subtraction reduces the noise components in a signal by estimating the level of noise in the
`
`signal. ('345 Patent, col. 1:58-60.) Prior techniques estimated the level of noise in the signal by
`
`measuring the magnitude of the signal "during non-speech time intervals detected by a voice
`
`switch" and using that estimate to subtract the noise from the signal. ('345 Patent, col. 1:60-64.)
`
`The problem with these prior art voice switches is that they had difficulty setting a threshold for
`
`the voice detector to accurately detect the non-speech intervals. ('345 Patent, col. 2:45-51.) A
`
`threshold set too high runs the "risk that some voice time intervals might be regarded as a non-
`
`speech time interval and the system will regard voice information as noise," resulting in
`
`distortion of the voice signal. ('345 Patent, col. 2:51-55.) A threshold set too low, however,
`
`runs the risk of erroneously detecting speech intervals during non-speech intervals, thus leading
`
`to non-speech intervals that are too short and producing poor estimates of the level of noise.
`
`('345 Patent, col. 2:55-58.)
`
`27.
`
`The '345 Patent addresses the short comings of the prior art noise estimation
`
`technique by determining non-speech segments by using a threshold detector for each frequency
`
`bin. ('345 Patent, col. 3:28-3 1.) In a preferred embodiment, "[t]he threshold detector precisely
`
`-16-
`
`IPR No. 2017-00626
`Apple Inc. v. Andrea Electronics Inc. - Ex. 1029, p. 16
`
`
`
`
`
`
`
`
`
`
`detects the positions of the noise elements, even within continuous speech segments, by
`
`determining whether frequency spectrum elements, or bins, of the input signal are within a
`
`threshold set according to a minimum value of the frequency spectrum elements over a preset
`
`period of time." ('345 Patent, col. 3:31-37.) This technique allows for a stable estimation of the
`
`noise level in a signal. The spectral subtraction technique described in the '345 Patent can be
`
`applied in various contexts, including "on an embedded hardware (DSP) as a stand-alone
`
`system" or "as a software application running on a PC using data obtained from a sound port."
`
`('345 Patent, col. 7:55-62.)
`
`28.
`
`The preferred embodiment of the spectral subtraction system of the '345 Patent is
`
`illustrated in Figure 1 of the '345 Patent, shown below:
`
`102
`
`Input
`Samples
`
`104
`
`106
`
`Collect
`Input
`Data
`
`Combine '
`I
`256 New
`J
`Point with
`1256 History J
`
`Shading
`Coefficients
`108
`
`I
`I
`l Hanning
`Multiply
`Window J
`
`512 Point
`FFT
`
`I
`[Noise
`j
`Processing
`
`IFFT
`I
`
`Overlap I
`I
`ii And
`[Sum J
`
`110
`
`112 (200)
`
`114
`
`116
`
`Output
`Samples
`
`118
`
`Figure 7: Figure 1 of the '345 Patent, which illustrates the preferred embodiment of the spectral
`subtraction system described in the '345 Patent.
`
`Digital input samples representing the input signal 102 are placed into a buffer. In particular, the
`
`preferred embodiment collects 256 digital input samples of the input signal at a time and places
`
`them into a temporary buffer 104. ('345 Patent, col. 4:65-66.) The 256 new digital input
`
`samples are then combined with the previous 256 points to provide 512 input points. ('345
`
`-17-
`
`IPR No. 2017-00626
`Apple Inc. v. Andrea Electronics Inc. - Ex. 1029, p. 17
`
`
`
`Patent, col. 4:66-5:1.) As such, the spectral subtraction system continually processes 256 new
`
`samples of the input audio signal 512 samples at a time. The 512 input points are multiplied by a
`
`Hanning Window 108 to "smooth the transients between two processed blocks" (i.e., the 256
`
`point blocks) and "prevent the masking of low energy tona!s by high energy side lobes." ('345
`
`Patent, col. 5:4-10.) The output of the Harming Window operation is then processed by a 512
`
`point Fast Fourier Transform (FFT) to generate 512 frequency components, or bins, of the input
`
`audio signal. ('345 Patent, col. 5:10-17.) In particular, the output of the FFT is a complex vector
`
`of 512 points, which are then processed by the noise processing block 112 (200). ('345 Patent,
`
`col. 5:17-21.) While the preferred embodiment describes the use of a FFT to generate the
`
`frequency bins of the input signal, the '345 Patent contemplates that "other transforms may be
`
`applied to the present invention to obtain the spectral noise signal." ('345 Patent, co!. 5:30-33.)
`
`29.
`
`Processing the input audio signal to generate its spectral content is a common task
`
`in signal analysis. This processing step is performed using a block of digital input samples of
`
`some length, where successive blocks typically overlap. Because both the digital input samples
`
`and the transformed output samples are indexed by time, this operation can be interpreted as a set
`
`of digital filters applied to the digital input samples to produce the transformed output samples.
`
`A general structure for this type of processing is called a filter bank. The most common
`
`transform is a weighted sum of the input samples, that is, a linear transform. An important type
`
`of linear transform is the Discrete Fourier Transform (DFT) also referred to as the Fast Fourier
`
`Transform which is an efficient implementation of the DFT. Other filter banks are also possible
`
`depending on the chosen processing. Each of these structures outputs signal components that are
`
`indexed by frequency and vary with time.
`
`IPR No. 2017-00626
`Apple Inc. v. Andrea Electronics Inc. - Ex. 1029, p. 18
`
`
`
`
`
`30.
`
`An embodiment of the noise estimation process is illustrated in Figure 2 of the
`
`'345 Patent. The complex vector generated by the FFT in Figure 1 is illustrated in item 202 in
`
`Figure 2.
`
`202
`>
`R(0)
`
`204
`
`206
`
`208
`
`R(n) 1(n)
`
`Y(n)=Max[R(n)l(n)J
`+04*Min[R(n)I(n)]
`
`1/31Y(n-1)+Y(n)+Y(n+ 1)]
`
`Y(n)t*0.3+Y(n)ti*07
`
`210
`
`i
`$
`j Subtraction
`l Process
`
`212(300)
`~
`
`$
`Noise
`Estimation
`
`Time Domain
`2141,,,' Input Signal
`
`Residual
`Process
`
`Output To
`IFFT
`
`218
`216
`Figure 8: Figure 2 of the '345 Patent, which illustrates the preferred embodiment of noise
`processing in the '345 Patent.
`
`In particular, R(0) and 1(0) in item 202 correspond to the real and imaginary components,
`
`respectively, of the input signal contained within the first frequency bin. The real and imaginary
`
`components of the input signal together represent the magnitude and phase components of the
`
`input signal within the frequency bin. Similarly, R(n) and 1(n) in item 202 correspond to the real
`
`and imaginary components of the input signal in a frequency bin with index n.
`
`31.
`
`Depending on the choice of processing, the frequency bins of the input signal may
`
`be represented as real-valued or complex-valued components. A linear transform that use