`571-272-7822
`
`Paper 34
`Entered: June 11, 2024
`
`UNITED STATES PATENT AND TRADEMARK OFFICE
`
`BEFORE THE PATENT TRIAL AND APPEAL BOARD
`
`APPLE INC., AMAZON WEB SERVICES, INC., and
`AMAZON.COM SERVICES LLC
`
`Petitioner,
`v.
`ZENTIAN LIMITED,
`Patent Owner.
`
`IPR2023-000371
`Patent 10,971,140 B2
`
`
`
`
`
`
`
`
`
`Before KEVIN F. TURNER, JEFFREY S. SMITH, and
`CHRISTOPHER L. OGDEN, Administrative Patent Judges.
`Opinion for the Board filed by Administrative Patent Judge SMITH.
`
`Opinion Concurring filed by Administrative Patent Judge OGDEN
`
`SMITH, Administrative Patent Judge.
`
`JUDGMENT
`Final Written Decision
`Determining No Challenged Claims Unpatentable
`35 U.S.C. § 318(a)
`
`
`
`
`
`
`1 IPR2023-01197 has been joined with this proceeding.
`
`
`
`
`
`IPR2023-00037
`Patent 10,971,140 B2
`
`INTRODUCTION
`I.
`A. Background and Summary
`Petitioner filed a Petition (Paper 1, “Pet.”) requesting inter partes
`review of claims 1–8 of U.S. Patent No. 10,971,140 B2 (Ex. 1001, “the
`’140 patent”). We issued an Institution Decision (Paper 10, “Dec.”)
`instituting the petitioned review. Patent Owner then filed a Patent Owner
`Response (Paper 19, “PO Resp.”) to the Petition. Petitioner filed a Reply
`(Paper 22, “Reply”) to the Patent Owner Response. Patent Owner filed a
`Sur-reply (Paper 27, “PO Sur-Reply”) to the Reply. An oral hearing was
`held on March 11, 2024, for which the transcript was entered into the record
`(Paper 33).
`We have jurisdiction under 35 U.S.C. § 6(b)(4) and § 318(a). This
`Decision is a final written decision under 35 U.S.C. § 318(a) and 37 C.F.R.
`§ 42.73 as to the patentability of claims 1–8 of the ’140 patent. We
`determine Petitioner has not shown by a preponderance of evidence that
`claims 1–8 are unpatentable.
`B. Related Matters
`The parties indicate that the following matters relate to the ’140
`patent: Zentian Ltd v. Apple Inc., 6:22-cv-00122 (W.D. Tex. Feb. 2, 2022);
`Zentian Ltd v. Amazon.com, Inc., 6:22-cv-00123 (W.D. Tex. Feb. 2, 2022);
`Apple Inc. v. Zentian Ltd., Inter Partes Review No. IPR2023-00033; Apple
`Inc. v. Zentian Ltd., Inter Partes Review No. IPR2023-00034; Apple Inc. v.
`Zentian Ltd., Inter Partes Review No. IPR2023-00035; and Apple Inc. v.
`Zentian Ltd., Inter Partes Review No. IPR2023-00036. Paper 4, 1; Pet. 64.
`
`2
`
`
`
`IPR2023-00037
`Patent 10,971,140 B2
`
`C. The ’140 Patent
`The ’140 patent is related to a speech recognition circuit which uses
`parallel processors for processing the input speech data in parallel.
`Ex. 1001, 1:18–20.
`The patent describes that in speech recognition, there are generally
`two processes: “front end processing to generate processed speech
`parameters such as feature vectors, followed by a search process which
`attempts to find the most likely set of words spoken from a given vocabulary
`(lexicon).” Id. at 1:21–26. According to the ’140 patent, “for large
`vocabulary, speaker independent speech recognition, it is the search process
`that presents the biggest challenge.” Id. at 1:28–30.
`The ’140 patent describes that in order to speed up the search
`function, parallel processing techniques have been suggested. Id. at 1:45–
`47. The patent further describes that “one algorithm for performing the
`search is the Viterbi algorithm,” which “is a parallel or breadth first search
`through a transition network of states of Hidden Markov Models.” Id. at
`1:36–39. This search algorithm is computationally intensive. Id. at 1:44. In
`one paper cited by the ’140 patent, “a multi-threaded implementation of a
`fast beam search algorithm is disclosed.” Id. at 1:47–52. This “multi-
`threading implementation requires a significant amount of communication
`and synchronization among threads.” Id. at 1:52–54. In another cited paper,
`“the parallel processing of input speech parameters is disclosed in which a
`lexical network is split statically among processors.” Id. at 1:54–58.
`To implement parallel processing of the search function, the ’140
`patent describes a special circuit, in which a “plurality of lexical tree
`processors are connected in parallel to the input port and perform parallel
`lexical tree processing for word recognition by accessing the lexical data in
`
`3
`
`
`
`IPR2023-00037
`Patent 10,971,140 B2
`the lexical memory arrangement.” Id. at 2:4–8. In addition, a “controller
`controls the lexical tree processors to process lexical trees identified in the
`results memory arrangement by performing parallel processing of a plurality
`of lexical tree data structures.” Id. at 2:12–15.
`Figure 2 is a diagram of the circuit of the ’140 patent, and is
`reproduced below.
`
`
`Figure 2, showing a plurality k of lexical tree processors 21, arranged
`in a lexical tree processor cluster 22, with acoustic model memory 23.
`D. Illustrative Claim
`Challenged claim 1 of the ’140 patent recites:
`1. [Pre] A speech recognition circuit comprising:
`[a] one or more clusters of processors, each of the one or more
`clusters of processors comprising:
`a plurality of processors; and
`[b] an acoustic model memory storing acoustic model
`data, [c] wherein each of the plurality of processors
`is configured to compute a probability using the
`
`4
`
`
`
`IPR2023-00037
`Patent 10,971,140 B2
`acoustic model data in the acoustic model memory,
`[d] wherein:
`the speech recognition circuit is configured to
`generate an initial score for an audio sample;
`and
`[e] the initial score is used to determine whether to
`continue processing to determine a final
`score via processing a larger amount of
`model data than that was processed to
`generate the initial score.
`Ex. 1001, 12:13–26; Pet. 66–67 (showing Petitioner’s bracketed claim
`annotations).
`
`E. Evidence
`Petitioner relies on the following prior art:
`U.S. Patent No. 6,374,219 B1, issued April 16, 2002 (Ex. 1004,
`“Jiang”);
`U.S. Patent No. 5,428,803, issued June 27, 1995 (Ex. 1005, “Chen”);
`U.S. Patent Appl. Publ. No. 2001/0053974 A1, published December
`20, 2001 (Ex. 1008, “Lucke”);
`U.S. Patent No. 5,983,180, issued November 9, 1999 (Ex. 1009,
`“Robinson”);
`U.S. Patent No. 5,036,539, issued July 30, 1991 (Ex. 1010,
`“Wrench”).
`
`F. Prior Art and Asserted Grounds
`Petitioner asserts that claims 1–8 of the ’140 patent are unpatentable
`on the following grounds:
`Claim(s) Challenged 35 U.S.C. §
`1–3, 5, 7, 8
`103(a)
`1–3, 5, 7, 8
`103(a)
`4
`103(a)
`
`Reference(s)/Basis
`Jiang, Chen
`Jiang, Chen, Lucke
`Jiang, Chen, Robinson
`
`5
`
`
`
`IPR2023-00037
`Patent 10,971,140 B2
`Claim(s) Challenged 35 U.S.C. §
`4
`103(a)
`6
`103(a)
`6
`103(a)
`
`Reference(s)/Basis
`Jiang, Chen, Lucke, Robinson
`Jiang, Chen, Wrench
`Jiang, Chen, Lucke, Wrench
`
`Pet. 6. Petitioner relies on the Declaration testimony of Christopher
`Schmandt. Ex. 1003.
`
`II. ANALYSIS
`A. Legal Standards
`“In an [inter partes review], the petitioner has the burden from the
`onset to show with particularity why the patent it challenges is
`unpatentable.” Harmonic Inc. v. Avid Tech., Inc., 815 F.3d 1356, 1363
`(Fed. Cir. 2016) (citing 35 U.S.C. § 312(a)(3) (requiring inter partes review
`petitions to identify “with particularity . . . the evidence that supports the
`grounds for the challenge to each claim”)); see also 37 C.F.R. § 42.104(b)
`(requiring a petition for inter partes review to identify how the challenged
`claim is to be construed and where each element of the claim is found in the
`prior art patents or printed publications relied upon).
`A claim is unpatentable under 35 U.S.C. § 103(a) if “the differences
`between the subject matter sought to be patented and the prior art are such
`that the subject matter as a whole would have been obvious at the time the
`invention was made to a person having ordinary skill in the art to which said
`subject matter pertains.” KSR Int’l Co. v. Teleflex Inc., 550 U.S. 398, 406
`(2007). The question of obviousness is resolved on the basis of underlying
`factual determinations, including: (1) the scope and content of the prior art;
`(2) any differences between the claimed subject matter and the prior art;
`(3) the level of skill in the art; and (4) when in evidence, objective evidence
`of obviousness or nonobviousness, i.e., secondary considerations. See
`
`6
`
`
`
`IPR2023-00037
`Patent 10,971,140 B2
`Graham v. John Deere Co., 383 U.S. 1, 17–18 (1966). An obviousness
`analysis “need not seek out precise teachings directed to the specific subject
`matter of the challenged claim, for a court can take account of the inferences
`and creative steps that a person of ordinary skill in the art would employ.”
`KSR, 550 U.S. at 418.
`Additionally, the obviousness inquiry typically requires an analysis of
`“whether there was an apparent reason to combine the known elements in
`the fashion claimed by the patent at issue.” KSR, 550 U.S. at 418 (citing
`In re Kahn, 441 F.3d 977, 988 (Fed. Cir. 2016) (requiring the fact-finder to
`provide “some articulated reasoning with some rational underpinning to
`support the legal conclusion of obviousness”)). Furthermore, Petitioner does
`not satisfy its burden of proving obviousness by employing “mere
`conclusory statements,” but “must instead articulate specific reasoning,
`based on evidence of record, to support the legal conclusion of
`obviousness.” In re Magnum Oil Tools Int’l, Ltd., 829 F.3d 1364, 1380
`(Fed. Cir. 2016).
`
`B. Claim Construction
`We construe the challenged claims
`using the same claim construction standard that would be used to
`construe the claim in a civil action under 35 U.S.C. 282(b),
`including construing the claim in accordance with the ordinary
`and customary meaning of such claim as understood by one of
`ordinary skill in the art and the prosecution history pertaining to
`the patent.
`37 C.F.R. § 42.100(b) (2022). “In determining the meaning of [a] disputed
`claim limitation, we look principally to the intrinsic evidence of record,
`examining the claim language itself, the written description, and the
`prosecution history, if in evidence.” DePuy Spine, Inc. v. Medtronic
`
`7
`
`
`
`IPR2023-00037
`Patent 10,971,140 B2
`Sofamor Danek, Inc., 469 F.3d 1005, 1014 (Fed. Cir. 2006). Claim terms
`are given their ordinary and customary meaning as would be understood by
`a person of ordinary skill in the art at the time of the invention and in the
`context of the entire patent disclosure. Phillips v. AWH Corp., 415 F.3d
`1303, 1313 (Fed. Cir. 2005) (en banc). “There are only two exceptions to
`this general rule: 1) when a patentee sets out a definition and acts as his own
`lexicographer, or 2) when the patentee disavows the full scope of a claim
`term either in the specification or during prosecution.” Thorner v. Sony
`Comput. Entm’t Am. LLC, 669 F.3d 1362, 1365 (Fed. Cir. 2012).
`Petitioner “applies the plain and ordinary meaning of all claim terms
`[as understood by a person of ordinary skill in the art for all terms].” Pet. 7.
`Patent Owner does not offer an explicit claim construction. See generally
`PO Resp. Although neither party explicitly construes the claims, the parties
`implicitly construe the claims in their respective arguments over level of
`ordinary skill and patentability. We agree with Petitioner that the proper
`standard for construing the claims is the plain and ordinary meaning
`standard. To the extent necessary, we resolve the implicit claim construction
`disputes by applying the plain and ordinary meaning to the claims, as
`discussed below.
`
`C. Level of Ordinary Skill in the Art
`Petitioner contends that a person having ordinary skill in the art would
`have had “a master’s degree in computer engineering, computer science,
`electrical engineering, or a related field, with at least two years of experience
`in the field of speech recognition, or a bachelor’s degree in the same fields
`with at least four years of experience in the field of speech recognition,” and
`that “[a]dditional education or experience might substitute for the above
`requirements.” Pet. 5 (citing Ex. 1003 ¶¶ 24–26).
`
`8
`
`
`
`IPR2023-00037
`Patent 10,971,140 B2
`Patent Owner contends that the field of electrical engineering has
`many specialties, such as digital signal processing and computer
`architecture. PO Resp. 6 (citing Ex. 2020 ¶ 17). Patent Owner contends that
`a person of ordinary skill in the field of speech recognition would pursue
`studies specific to digital signal processing, and a person of ordinary skill in
`the field of computer architecture would take a different set of courses. Id.
`Patent Owner contends that a person of ordinary skill in speech recognition
`would not be expected to have also specialized in parallel processing
`architectures and methods, or in high performance computing. Id. at 7
`(citing Ex. 2020 ¶ 18). Patent Owner contends that Petitioner’s declarant,
`Mr. Schmandt, has not taught a course that would have equipped a masters-
`level student in speech recognition to apply a parallel processing architecture
`to known speech recognition techniques. Id. Patent Owner contends that
`although a person with a master’s degree in one of the fields identified by
`Mr. Schmandt could be a person of ordinary skill in speech recognition, or a
`person of ordinary skill in high performance computing and parallel
`processing, that person would not be both. Id. (citing Ex. 2020 ¶ 19).
`We do not need to resolve this dispute, because even under
`Petitioner’s definition of the level of ordinary skill, Petitioner has not shown
`that the challenged claims are unpatentable. We apply Petitioner’s definition
`that a person having ordinary skill in the art would have had a master’s
`degree in computer engineering, computer science, electrical engineering, or
`a related field, with two years of experience in the field of speech
`recognition, or a bachelor’s degree in the same fields with four years of
`experience in the field of speech recognition, and that additional education
`or experience might substitute for the above requirements.
`
`9
`
`
`
`IPR2023-00037
`Patent 10,971,140 B2
`D. Obviousness over Jiang and Chen
`Petitioner contends that claims 1–3, 5, 7, and 8 would have been
`obvious over the combined teachings of Jiang and Chen. Pet. 9–53.
`1. Overview of Jiang
`Jiang, entitled “System for Using Silence in Speech Recognition,” is
`related to “computer speech recognition performed by conducting a prefix
`tree search of a silence bracketed lexicon.” Ex. 1004, code (54), 1:16–18.
`Figure 2 shows a block diagram of speech recognition system 60, and
`is reproduced below.
`
`
`
`Figure 2 above shows microphone 62, analog-to-digital (A/D)
`converter 64, training module 65, feature extraction module 66, silence
`detection module 68, lexicon storage module 70, phonetic speech unit
`storage module 72, tree search engine 74, and output device 76. Id. at 6:20–
`25.
`
`Jiang discloses that the speech recognition system:
`recognizes speech based on an input data stream indicative of the
`speech. Possible words represented by the input data stream are
`provided as a prefix tree including a plurality of phoneme
`branches connected at nodes. The plurality of phoneme branches
`
`10
`
`
`
`IPR2023-00037
`Patent 10,971,140 B2
`are bracketed by at least one input silence branch corresponding
`to a silence phone on an input side of the prefix tree and at least
`one output silence branch corresponding to a silence phone on an
`output side of the prefix tree.
`Id. at 4:7–16. Lexicon storage module 70 contains information, which is
`representative of all of the words in the vocabulary of speech recognition
`system 60. Id. at 7:65–67. The words are presented to tree search engine 74
`in the form of a prefix tree, which can be traversed from a root to a leaf to
`arrive at the word most likely indicative of the utterance of the user. Id. at
`7:67–8:4. As the tree is traversed from an input node to an output node, a
`score is assigned to each node connected to a phoneme branch then under
`consideration. Id. at 8:16–20.
`A pruning technique can be used by comparing the score at a given
`node with the largest score from the other nodes of a frame being
`considered. Id. at 8:52–55. If the score at that node is sufficiently lower
`than the largest score, that branch is pruned from the tree, thereby drastically
`reducing the search space. Id. at 8:55–64.
`Overview of Chen
`2.
`Chen “relates generally to parallel processing computer systems for
`performing multiple-instruction-multiple-data (MIMD) parallel processing.”
`Ex. 1005, 1:29–31. Chen describes an architecture for high performance
`MIMD multiprocessors, which organizes the multiprocessors into four or
`more physically separable clusters, and provides for a shared memory model
`to be used with programs executed in the floating shared memory space, and
`a distributed memory model to be used with any programs executed across
`non-adjacently interconnected clusters. Id. at 1:32–43.
`Figure 4 is a block diagram of the architecture of the Chen system,
`annotated by Petitioner, and is reproduced below.
`
`11
`
`
`
`IPR2023-00037
`Patent 10,971,140 B2
`
`
`Figure 4 above shows computer clusters 100a (in red), 100b (in green), 100c
`(in blue), and 100d (in green), processors (P) 102a–d, and memories 104a–d.
`Chen’s architecture provides all three of the major memory types for parallel
`processing:
`A true shared memory model with symmetrical access to a
`common shared memory 104a is provided for all processors
`102a in cluster 100a. An extended shared memory model is
`provided for all of the processors 102a, 102b and 102c that
`adjacently access the cluster shared memories 104a, 104b and
`104c in the floating shared memory 110, for example. Finally, a
`distributed shared memory model is provided for all processors
`102a that need to access the cluster shared memory 104d of a
`non-adjacently connected cluster 100d, for example.
`
`12
`
`
`
`IPR2023-00037
`Patent 10,971,140 B2
`Id. at 9:44–56.
`
`3. Claim 1
`Claim 1(a): “one or more clusters of processors . . . .”
`Claim 1 recites “one or more clusters of processors, each of the one or
`more clusters of processors comprising: a plurality of processors.”
`Petitioner contends that the combination of Jiang and Chen teaches
`this limitation. Petitioner contends that the “’140 Patent describes a ‘cluster’
`as at least including processing architectures comprising a group of
`processors and a memory.’” Pet. 14 (citing Ex. 1001, 3:14–19, 3:53–57,
`5:14–21, Fig. 2).
`Petitioner contends that Jiang teaches implementing its speech
`recognition system on multiprocessor systems, but does not specifically
`discuss clusters of processors. Id. (citing Ex. 1001, 4:51–5:3, 6:39–42).
`Petitioner contends that Chen teaches clusters of processors, with each
`cluster comprising a plurality of processors and memory. Id. (citing Ex.
`1005, 9:5–43). In particular, Petitioner contends that Figure 4 of Chen
`shows four clusters, 100a, 100b, 100c, and 100d, that together comprise a
`parallel processing computer system. Petitioner contends that each cluster
`such as 100a includes two or more processors 102a that are symmetrically
`connected to cluster shared memory 104a via connection node 106a. Pet. 15
`(citing Ex. 1004, 9:5–19, Fig. 4; Ex. 1003 ¶ 73); see id. at 16–17 (citing Ex.
`1004, Abstract, 5:9–17, 6:17–21, 9:5–43, 10:14–35, Figs. 4, 5A, 6A, 6B).
`Petitioner contends that Chen teaches “each of the one or more clusters of
`processors comprising: a plurality of processors” as claimed in disclosing
`clusters 100a, 100b, 100c, and 100d, each comprising respective processors
`102a, 102b, 102c, and 102d. Pet. 14.
`
`13
`
`
`
`IPR2023-00037
`Patent 10,971,140 B2
`Petitioner contends that a person of ordinary skill in the art would
`have replaced Jiang’s tree search engine 74 with Chen’s clusters of
`processors in order to replicate the tree search engine’s functionality
`amongst each of Chen’s processors for the benefit of processing hidden
`Markov models in parallel. Pet. 17–19 (citing Ex. 1003 ¶¶ 71–72, 74).
`Patent Owner disputes Petitioner’s reason to combine Jiang and Chen.
`PO Resp. 17–28. We agree with Patent Owner for the reasons given below
`in our analysis of reasons to combine Jiang and Chen.
`Claim 1(b): “an acoustic model memory storing acoustic model data”
`Claim 1 recites “each of the one or more clusters of processors
`comprising . . . an acoustic model memory storing acoustic model data.”
`Petitioner contends that Jiang’s phonetic speech unit model memory
`72 stores phonetic speech unit models such as hidden Markov models that
`represent phonemes. Pet. 19–22 (citing Ex. 1004, Fig. 2, 6:18–28, 7:29–54).
`Petitioner, relying on the testimony of Mr. Schmandt, contends that a person
`of ordinary skill in the art would have recognized that Jiang’s hidden
`Markov model (HMM) based phonetic model teaches the claimed acoustic
`model. Pet. 21 (citing Ex. 1003 ¶ 75). Petitioner contends that Jiang’s
`phonetic speech unit model memory 72 therefore teaches “an acoustic model
`memory storing acoustic model data” as claimed. Pet. 22. Petitioner
`contends that Chen teaches “one or more clusters of processors,” where each
`cluster includes a cluster shared memory. Pet. 22–24 (citing Ex. 1005, Fig.
`4, Abstract, 5:9–17, 6:17–21, 9:5–43, 10:14–35; Ex. 1003 ¶ 76).
`Petitioner, relying on the testimony of Mr. Schmandt, contends that a
`person of ordinary skill in the art would have stored at least a portion of
`Jiang’s acoustic model data in the cluster shared memory of each cluster so
`that each processor would have access to the necessary acoustic model data
`
`14
`
`
`
`IPR2023-00037
`Patent 10,971,140 B2
`for performing speech recognition for the benefits of reducing the time
`required to recognize a user’s speech, enabling a more relaxed pruning
`threshold, and improving accuracy. Pet. 24–25 (citing Ex. 1003 ¶¶ 72, 74,
`76).
`
`Patent Owner disputes Petitioner’s reason to combine Jiang and Chen.
`PO Resp. 17–28. We agree with Patent Owner for the reasons given below
`in our analysis of the reasons to combine Jiang and Chen.
`Claim 1(c): “wherein each of the plurality of processors . . . .”
`Claim 1 recites “wherein each of the plurality of processors is
`configured to compute a probability using the acoustic model data in the
`acoustic model memory.”
`Petitioner contends that Jiang teaches a processor “configured to
`compute a probability using the acoustic model data” in disclosing that tree
`search engine 74 determines a most likely phoneme represented by a
`codeword based upon hidden Markov models stored in memory 72. Pet. 25–
`30 (citing Ex. 1004, 1:28–32, 1:33–39, 2:1–15, 7:30–45, 8:16–52; Ex. 1003
`¶¶ 77–82). Petitioner contends that Chen teaches that for each cluster of
`processors, the processors in each cluster may perform processing in
`parallel, such that each processor of each of the clusters performs the
`intended operation. Pet. 30 (citing Ex. 1005, 5:10–17, 18:64–19:9; Ex. 1003
`¶ 83). Petitioner contends that in the system of Jiang as modified by Chen,
`each of Chen’s processors 102a through 102d would have computed
`probability scores as taught by Jiang using Jiang’s acoustic model data
`stored in shared cluster memories 104a through 104d. Id.
`Petitioner contends that a person of ordinary skill would have
`modified Jiang to include Chen’s clusters of processors and would have
`caused each processor to perform a lexical tree search using acoustic model
`
`15
`
`
`
`IPR2023-00037
`Patent 10,971,140 B2
`data for the reasons given in Petitioner’s analysis of limitations 1(a) and
`1(b). Pet. 31.
`Patent Owner disputes Petitioner’s reason to combine Jiang and Chen.
`PO Resp. 17–28. We agree with Patent Owner for the reasons given below
`in our analysis of the reasons to combine Jiang and Chen.
`Claim 1(d): “wherein: the speech recognition circuit is configured . . .” and
`claim 1(e): “the initial score is used to determine . . . .”
`Claim 1 recites “wherein: the speech recognition circuit is configured
`to generate an initial score for an audio sample.” Claim 1 further recites “the
`initial score is used to determine whether to continue processing to
`determine a final score via processing a larger amount of model data than
`that was processed to generate the initial score.”
`Summary of contentions regarding limitations 1(d) and 1(e)
`Petitioner contends that Jiang teaches an audio sample in disclosing
`that a microphone outputs an acoustic signal that is processed to generate a
`codeword, where the codeword corresponds to the “audio sample” as
`claimed. Pet. 31–32 (citing Ex. 1004, 1:58–67, 2:1–15, 6:46–7:1, 7:5–15,
`7:29–47, 8:16–24, 10:1–11; Ex. 1003 ¶¶ 78–82, 85, 87). Petitioner contends
`that Jiang teaches the claimed “speech recognition circuit is configured to
`generate an initial score” in disclosing a tree search engine that, in traversing
`prefix tree 77 shown in Figure 3, computes scores for nodes 82, 84, and 86,
`which are the nodes associated with the initial branches of the prefix tree
`leaving root node 78, such that the score for each of nodes 82, 84, and 86 is
`an “initial score” representing the likelihood that the phoneme represented
`by the first codeword under consideration corresponds to the phoneme
`represented by the branch associated with the node. Pet. 32–36 (citing Ex.
`
`16
`
`
`
`IPR2023-00037
`Patent 10,971,140 B2
`1004, 1:19–33, 2:1–15, 7:29–54, Ex. 1004, 8:5–51, Fig. 3; Ex. 1003 ¶¶ 84,
`87).
`
`Petitioner contends that Jiang teaches “the initial score is used to
`determine whether to continue processing to determine a final score via
`processing a larger amount of model data than was processed to generate the
`initial score” as claimed in disclosing a pruning technique that compares, at
`each node, the scores assigned to that node with the largest score on any of
`the other nodes corresponding to the frame being considered, and, if the
`score at a particular node is sufficiently low compared to the largest score,
`then pruning the corresponding branch from the prefix tree and no longer
`considering the pruned branch in further processing. Pet. 38–39 (citing Ex.
`1004, 8:52–64, 10:51–11:24); see id. at 39–44. Petitioner, relying on
`testimony of Mr. Schmandt, contends that because the initial score is used to
`determine whether to continue processing to determine a final score or to
`prune the branch, the final score is determined by processing a larger amount
`of model data than was processed to generate the initial score. Pet. 42–44
`(citing Ex. 1003 ¶¶ 90–91).
`In its Response, Patent Owner contends that neither the Petition nor
`Mr. Schmandt’s Declaration explains how Jiang’s tree search engine
`modified in view of Chen would operate to meet limitations 1(d) and 1(e).
`PO Resp. 29–32 (citing Ex. 2017, 79:11–23, 81:8–22, 83:15–18). Patent
`Owner contends that Mr. Schmandt testified during cross-examination that
`one processor in one cluster in Chen would have been used to arrive at
`limitations 1(d) and 1(e). Id. at 32 (citing Ex. 2017, 102:5–103:3 (Mr.
`Schmandt testifying that after the processors finish computing their
`respective scores by operating on the acoustic data in parallel, one processor
`can look at all of the scores and perform pruning)); see Ex. 2017, 45:10–
`
`17
`
`
`
`IPR2023-00037
`Patent 10,971,140 B2
`46:10, 97:8–17, 98:11–99:7, 101:10–23. Patent Owner contends Mr.
`Schmandt is incorrect, because Chen teaches that not all the clusters are
`interconnected, and that processors 102a in cluster 100a cannot directly or
`adjacently access memory 104d of cluster 100d, while processors 102d
`cannot directly or adjacently access memory 104a. PO Resp. 33 (citing
`Ex. 1003 ¶ 95). According to Patent Owner, because the scores for the
`processors in each cluster are stored in that cluster’s memory, and because
`no one processor could access all of the scores stored in all of the memories,
`no one processor could perform “the initial score is used to determine
`whether to continue processing” the model data as claimed. Id. at 35 (citing
`Ex. 1005, 9:10–39; Ex. 2020 ¶ 51).
`In Reply, Petitioner contends that the Petition states that “Jiang’s
`algorithm of generating an initial score and using the initial score to
`determine whether to continue processing (the functionality of Claims 1(d)–
`1(e)) would be replicated across Chen’s processors.” Reply 2 (citing Pet.
`17–18, 24, 30; Ex. 1003, 74, 76–77, 83). Petitioner contends that the claim
`does not require any particular processor from the plurality of processors to
`perform any part of the functionality of limitations 1(d) and 1(e). Id. at 2–3.
`Petitioner contends that Mr. Schmandt confirmed during cross-examination
`that although paragraphs 83 through 87 of his Declaration did not explain
`how the combination of Jiang and Chen operated, other portions of his
`Declaration did discuss the combination. Id. at 2 (citing Ex. 2017, 44:2–11,
`45:10–46:10, 71:6–23, 73:12–17, 79:11–23, 81:8–22, 83:15–18, 86:25–
`87:23, 93:7–94:10, 94:16–23, 95:19–96:10, 97:8–17, 99:17–102:12).
`Petitioner contends that in the combination of Jiang and Chen, each
`processor operates on a portion of the lexical trees, which does not require
`inter-cluster communication. Id. at 16. Petitioner contends that once the
`
`18
`
`
`
`IPR2023-00037
`Patent 10,971,140 B2
`scores are calculated, inter-cluster processing of the results would be
`communicated. Id. Petitioner contends that the combination of Jiang and
`Chen as proposed in the Petition provides the framework for inter-cluster
`communication of the results. Id. at 17–18.
`In its Sur-Reply, Patent Owner contends that Petitioner’s argument in
`its Reply, that “Jiang’s algorithm of generating an initial score and using the
`initial score to determine whether to continue processing (the functionality
`of claims 1(d)–1(e)) would be replicated across Chen’s processors,” fails for
`several reasons. PO Sur-Reply 3 (quoting Reply 2). Patent Owner contends
`that Petitioner’s theory in its Reply requires every processor in Chen’s four
`clusters to perform limitations 1(d) and 1(e), in contrast to Mr. Schmandt’s
`deposition theory, which required only one processor to do so, thus
`abandoning Mr. Schmandt’s deposition theory and discrediting Mr.
`Schmandt. Id. Patent Owner contends that Mr. Schmandt testified at his
`deposition that there would be no advantage from using all of Chen’s
`processors to calculate the initial score and determine whether to continue
`processing because “[a]t that point, we have finished our parallel processing
`operation and we’re back to serial operation.” Id. at 3–4 (quoting Ex. 2017,
`102:5–103:3).
`Patent Owner contends that Petitioner’s theory would require every
`processor in Chen’s four clusters to calculate the same initial score and make
`the same determination whether to continue processing towards a final score,
`which makes no sense. PO Sur-Reply 3. Patent Owner contends that
`Petitioner’s theory in its Reply requires each processor to obtain necessary
`information stored across all of Chen’s four memories 104a through 104d,
`even though none of Chen’s processors can directly or adjacently access all
`cluster memories. Id. at 4. Patent Owner contends that Petitioner’s theory
`
`19
`
`
`
`IPR2023-00037
`Patent 10,971,140 B2
`in Reply was not set forth in the Petition, and that Petitioner’s citations are to
`aspects of the Petition discussing other limitations, not limitations 1(d) and
`1(e). Id.
`Analysis of the contentions regarding limitations 1(d) and 1(e)
`We agree with Patent Owner that the Petition’s combination of Jiang
`and Chen requires implementing Jiang’s algorithm on a plurality of clusters
`of processors as shown in Figure 4 of Chen. PO Sur-Reply 27; Ex. 2020
`¶ 44; see Pet. 15–17, 19–24; Ex. 1003 ¶¶ 73–76, 83. We highlight that Chen
`discloses that its system includes at least four clusters of processors. Ex.
`1005, 9:13–14 (“Four clusters 100a, 100b, 100c and 100d together comprise
`a parallel processing computer system.”); see id. at 11:51–53 (“In order to
`provide a common shared memory, there should be four or more groups of
`processors.”).
`Even though Petitioner’s combination requires implementing Jiang’s
`algorithm on Chen’s plurality of clusters of processors, the Petition contends
`that Jiang alone teaches limitations 1(d) and 1(e) in disclosing assigning an
`initial score to each of the initial nodes of a prefix tree, then using these
`initial scores to determine, via pruning, whether to continue processing
`branches connected to the nodes. Pet. 38 (citi