Trials@uspto.gov
`571-272-7822
`
`Paper No. 33
`
`UNITED STATES PATENT AND TRADEMARK OFFICE
`__________
`
`BEFORE THE PATENT TRIAL AND APPEAL BOARD
`_______
`
`APPLE INC.,
`Petitioner,
`
`v.
`
`ZENTIAN LIMITED,
`Patent Owner.
`__________
`
`IPR2023-00037
`Patent 10,971,140 B2
`__________
`
`Record of Oral Hearing
`Held: March 11, 2024
`
`__________
`
`
`
`
`Before KEVIN F. TURNER, JEFFREY S. SMITH, and
`CHRISTOPHER L. OGDEN, Administrative Patent Judges.
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`

`

`IPR2023-00037
`Patent 10,971,140 B2
`
`
`
`
`
`APPEARANCES:
`
`ON BEHALF OF THE PETITIONER:
`
`JENNIFER BAILEY, ESQ.
`Erise IP, P.A.
`7015 College Boulevard
`Suite 700
`Overland Park, KS 66211
`(913) 777-5600
`Jennifer.bailey@eriseip.com
`CHRISTINA CANINO, ESQ.
`Erise IP, P.A.
`5299 DTC Boulevard
`Suite 1340
`Greenwood Village, CO 80111
`(720) 307-8382
`Christina.Canino@eriseip.com
`
`JENNY LIU, ESQ.
`Apple Inc.
`1 Apple Park Way
`Cupertino, CA 95014
`(408) 996-1010
`jianing_liu@apple.com
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`2
`
`
`
`

`

`IPR2023-00037
`Patent 10,971,140 B2
`
`ON BEHALF OF THE PATENT OWNER:
`
`KAYVAN NOROOZI, ESQ.
`PETER KNOPS, ESQ.
`Noroozi PC
`11601 Wilshire Boulevard
`Suite 2170
`Los Angeles, CA 90025
`(310) 975-7074 ex. 1
`kayvan@noroozipc.com
`peter@noroozipc.com
`
`JESSICA R. BERNHARDT, ESQ.
`Bartlit Beck LLP
`Courthouse Place
`54 West Hubbard Street
`Chicago, IL 60654
`(312) 494-4460
`Jessica.bernhardt@bartlitbeck.com
`
`DARGAYE CHURNET, ESQ.
`Fenwick & West LLP
`555 California St. #12
`San Francisco, CA 94104
`(415) 875-2460
`dchurnet@fenwick.com
`
`
`
`The above-entitled matter came on for hearing Monday, March 11,
`
`2023, commencing at 2:44 p.m. EDT, via Video-conference.
`
`
`
`
`
`3
`
`
`
`

`

`IPR2023-00037
`Patent 10,971,140 B2
`
`P-R-O-C-E-E-D-I-N-G-S
`
`2:44 p.m.
`
`COURT REPORTER: On the record.
`
`
`JUDGE OGDEN: Good afternoon. Welcome to the Patent
`
`
`Trial and Appeal Board. This is the hearing for IPR2023-00037, between
`Petitioner Apple Inc., Amazon Web Services, Inc., and Amazon.com
`Services LLC, and Patent Owner Zentian Limited, challenging claims in
`Patent 10,839,798 B2.
`
`
`And I'm Judge Ogden, and with me today on the panel are
`Judges Turner and Smith. Judge Smith is joining us remotely.
`
`
`So I'd like to go through counsel introductions as we did with
`the prior case this morning. So who is here on behalf of Petitioner Apple?
`
`
`MS. BAILEY: Thank you, Your Honor. I'm Jennifer Bailey.
`I'm from the law firm of Erise IP. I have with me my co-counsel, Christina
`Canino, and in-house counsel for Apple, Jenny Liu.
`
`
`JUDGE OGDEN: Thank you. And on behalf of Patent
`Owner?
`MR. NOROOZI: Yes, Your Honor. Your Honor, Kayvan
`
`
`Noroozi from Noroozi PC for Patent Owner, and with me are Peter Knops
`and Jessica Bernhardt.
`
`
`JUDGE OGDEN: And on behalf of the Amazon Petitioners?
`
`
`MR. CHURNET: Hello, Your Honor. Dargaye Churnet from
`Fenwick & West on behalf of Amazon.
`
`
`JUDGE OGDEN: Thank you. So today, each side will have a
`total of 45 minutes to present their arguments, and as usual, parties will have
`the opportunity to reserve rebuttal or surrebuttal time.
`
`1
`2
`3
`4
`5
`6
`7
`8
`9
`10
`11
`12
`13
`14
`15
`16
`17
`18
`19
`20
`21
`22
`23
`24
`25
`26
`
`
`
`
`
`4
`
`
`
`

`

`1
`2
`3
`4
`5
`6
`7
`8
`9
`10
`11
`12
`13
`14
`15
`16
`17
`18
`19
`20
`21
`22
`23
`24
`25
`26
`27
`
`IPR2023-00037
`Patent 10,971,140 B2
`
`
`So we'll begin with Petitioner. Ms. Bailey, would you like to
`reserve rebuttal time?
`
`
`MS. BAILEY: Yes, Your Honor. I'd like to reserve 15 minutes
`for rebuttal.
`JUDGE OGDEN: Okay. We'll put 30 minutes on the clock,
`
`
`and you can begin when you're ready.
`
`
`MS. BAILEY: May it please the board. Thank you, Your
`Honors.
`Let's turn first to Petitioner's Demonstrative DX-2, and let's
`
`
`quickly go through the claim. Claim 1 recites a basic distance calculation
`and search steps using a clustered processor architecture. The basic tree
`search steps are limitations, 1(d) and 1(e). And I note for limitations 1(d)
`and 1(e), it says the speech recognition circuit is configured to generate both
`the initial and final scores. The claim does not require any particular
`processor to generate the scores, but we'll get into that in a little bit.
`
`
`The claim ends with the tree search score generation. I know
`that in Claim 1, word recognition is not claimed, what happens to the scores
`after generation is not claimed, how the final scores are compiled, how the
`pruning is instructed, and how the word is selected are all not claimed.
`
`
`Turning to DX-3, let's go briefly through the combination.
`Jiang teaches a tree search algorithm that includes distance calculations and
`tree search scoring. Jiang's tree search algorithm is per the Petition's
`mapping, replicated across all of the processors of the clusters of Chen's
`clustered processor architecture. And I'm happy to provide citations as I go
`if that would be helpful. So each processor in each cluster is performing the
`lexical tree search recited in Claims 1(d) and 1(e). In other words, the initial
`and final score generation.
`
`
`
`
`
`5
`
`
`
`

`

`1
`2
`3
`4
`5
`6
`7
`8
`9
`10
`11
`12
`13
`14
`15
`16
`17
`18
`19
`20
`21
`22
`23
`24
`25
`26
`27
`
`IPR2023-00037
`Patent 10,971,140 B2
`
`
`And then the petition describes that the acoustic model is
`divided up among the memories of the claims. This means that a portion of
`the acoustic model data is stored in each memory's clusters. A portion is.
`So for example, Cluster 1 would perform tree search on phonemes A
`through F, Cluster 2 would be G through J. And he was questioned about it
`during his deposition, Mr. Schmandt, at Exhibit 2017, Pages 93 through 96,
`and 99 through 100. But also how the acoustic model is divided among each
`cluster's memory is also not claimed.
`
`
`So turning to DX-4. As mentioned, Claims 1(d) through 1(e)
`recite the initial and final score generation, and the mapping relies on Jiang
`for teaching the initial, final, and score generation. Zentian does not dispute
`that Jiang teaches the initial and final score generation. Instead, Zentian
`argues that no one processor in any of the Chen clusters can access all of the
`data stored across all of the clusters' memories. That's the quotation we have
`at the top of DX-4.
`
`
`And just so that we all have a clear understanding of Chen,
`Chen has an example of four clusters, and in Chen, each of the processors
`can access the memories of three other clusters. So technically, Zentian is
`correct in their statement, but the problem comes in is that is not required in
`the combination, and it's not required in the claims. There is nothing in the
`proposed combination that requires one processor from one cluster to be able
`to access all of the memories of the other clusters.
`
`
`Zentian argues all four memories contain necessary information
`for the score determination. And again, this is incorrect. Each processor is
`performing the score determination using the acoustic model data stored in
`that cluster's memory. And I really want to emphasize that because Zentian
`has completely revised and fabricated what the petition's mapping is.
`
`
`
`
`6
`
`
`
`

`

`1
`2
`3
`4
`5
`6
`7
`8
`9
`10
`11
`12
`13
`14
`15
`16
`17
`18
`19
`20
`21
`22
`23
`24
`25
`26
`27
`
`IPR2023-00037
`Patent 10,971,140 B2
`
`
`Jiang's tree search algorithm is replicated across all the
`processors so that each processor is performing the distance calculation, the
`probability, and the initial and final score determination. And we're going to
`get to it, but what happens to those scores is not claimed, is not recited in the
`claim.
`Jiang's tree search algorithm replicated on each processor does
`
`
`not require memory access between clusters. Zentian is bodily incorporating
`components from Chen into the combination that are neither required by the
`claims or required in the combination.
`
`
`Turning to DX-5. Zentian's sur-reply states that one processor
`and one cluster would be used to generate the initial score. Zentian
`concludes from that that one processor would have to access all of the
`memories of all of the clusters. And again, this sur-reply argument is a
`complete fabrication and is not what the petition mapped. The sur-reply
`arguments at Pages 3 to 4 are particularly incorrect.
`
`
`Turning to DX-6, Zentian's sole basis for this argument is a
`question posed to Mr. Schmandt and his answer. And the question that they
`rely on is the text on the right. I would invite the board to read kind of the
`lead up to this questioning to understand why Counsel asked a poor
`question, and what Mr. Schmandt was answering is not what Counsel has
`turned that answer into.
`
`
`And I refer to Exhibit 2017, Page 99, Line 8 through Page 103,
`Line 9. Looking at the left-hand side of the text on DX-6, Zentian's Counsel
`asked about the results of the tree search. That's the first highlight of
`section. And Mr. Schmandt said, well, we have to know the scores that we
`get on each of the trees. So remember, lexical tree search is performed, and
`initial and a final score are determined. Those scores are then used to apply
`
`
`
`
`7
`
`
`
`

`

`1
`2
`3
`4
`5
`6
`7
`8
`9
`10
`11
`12
`13
`14
`15
`16
`17
`18
`19
`20
`21
`22
`23
`24
`25
`26
`27
`
`IPR2023-00037
`Patent 10,971,140 B2
`the pruning threshold to prune the trees, but the pruning and how the scores
`are used is not claimed.
`
`
`Mr. Schmandt said that once the parallel operations were
`complete, in other words, once all the processors had completed their tree
`search and attained the results or the scores, then the system returns to serial
`operation. Because once we know the final score across all the trees, then
`the pruning can be performed. What will happen is one tree, oh, not highly
`likely prune. Another tree, it'll be likely it won't be pruned. But again, that's
`not claimed.
`
`
`The portion on the left shows that when Mr. Schmandt was
`asked about the results by Counsel, Mr. Schmandt was understanding the
`results to be the scores on the tree. Again, that's the second highlighted
`section on the left-hand side.
`
`
`And then Counsel asked, just so I'm clear, and I'm looking at
`the right-hand now, your testimony is that the result would be compiling or
`scored in one processor, for instance.
`
`
`Mr. Schmandt said, sure, you can take all the results and put
`them into one processor.
`
`
`But Mr. Schmandt was answering the question of what happens
`to all of the scores across all of the trees once the initial and final score
`determination is done. That's not claimed. So Mr. Schmandt was simply
`saying there's a way that you could do it. You could put all the scores, send
`them all to one processor, and then you could have one processor say prune,
`don't prune, prune, don't prune. But none of that is claimed in Claim 1.
`There's also probably other ways that it could be done.
`
`
`How the scores are compiled or aggregated from all the
`processors is not claimed. Mr. Schmandt was merely agreeing that one
`
`
`
`
`8
`
`
`
`

`

`1
`2
`3
`4
`5
`6
`7
`8
`9
`10
`11
`12
`13
`14
`15
`16
`17
`18
`19
`20
`21
`22
`23
`24
`25
`26
`27
`
`IPR2023-00037
`Patent 10,971,140 B2
`method of compiling the scores could be to send to one processor. But this
`is after all the parallel operations are done. In other words, after Jiang's tree
`search algorithm is done. So Zentian is here arguing unclaimed features and
`features that are not part of the claim combination.
`
`
`If there are no questions on this, I'm going to turn to the
`motivations to combine, DX-7. Zentian argues that there is a lack of
`reasonable expectations of success in the combination because there would
`be tasks to be performed that the skilled person would not know how to do.
`
`
`Zentian does not respond to the petition's arguable primary
`motivation that more processors in the combination allow a relaxed pruning
`threshold at the same financial cost. The motivation was that using Chen's
`commercially available processors in a clustered processor memory
`architecture then allows for a computing platform that's more powerful, has
`a more relaxed pruning threshold, which increases the speech recognition
`accuracy, and all of this is done at the same financial cause because you're
`using Chen's commercially available processors.
`
`
`Other motivations combined were discussed, including
`increasing the processing power while keeping the financial costs low for
`some of the same reasons I just discussed, the flexibility and scalability of
`using multiple clusters, which we know from home, and that clustered
`processors were a known technique. And I note to the board that a known
`technique alone is sufficient for a motivation to combine. That's the Intel v.
` Pact cited in our briefing. A motivation to combine is not needed for
`obviousness.
`
`
`I also want to address the surreplies briefing about Mathew.
`Zentian argues that Mathew shows that adding more processors actually
`slows down the system. The citation to Mathew is a single sentence from
`
`
`
`
`9
`
`
`
`

`

`IPR2023-00037
`Patent 10,971,140 B2
`Mathew that multiple processors were added to the HMM stage, and that
`multiple processors created synchronization issues.
`
`
`But again, this is a single sentence from Mathew. We have no
`indication in Mathew of what the architecture was when it added those
`additional processors. We have no indication that it was a clustered
`processor architecture. We have no discussion of what the division of the
`acoustic model was done. Mathew's single sentence teaching does not
`provide any context that's relevant to the motivation to combine inquiry
`here.
`Turning next to DX-8. And I want to pause. Are there any
`
`
`questions so far, Your Honors? No?
`
`
`JUDGE SMITH: I want to ask you about the claim.
`
`
`MS. BAILEY: Okay.
`
`
`JUDGE SMITH: The claim, what you're calling Claim 1(a),
`one or more clusters of processors, each of the one or more clusters of
`processors comprising a plurality of processors. And then nobody really
`construed that term, but the literal language of this term seems to say that
`one cluster of processors, which is a plurality of two processors, would be
`enough to meet this claim in 1(b), one model memory. So you could have
`one cluster of processors comprising two processors hooked up to one
`memory.
`Would you say that's accurate?
`
`
`MS. BAILEY: I'm actually smiling here, Your Honor, because
`
`
`I had the same debate when we were first putting together the petition. The
`one or more is properly interpreted as one cluster. And a cluster has to have
`a plurality of processors. A plurality would be understood to be two or
`
`1
`2
`3
`4
`5
`6
`7
`8
`9
`10
`11
`12
`13
`14
`15
`16
`17
`18
`19
`20
`21
`22
`23
`24
`25
`26
`
`
`
`
`
`10
`
`
`
`

`

`1
`2
`3
`4
`5
`6
`7
`8
`9
`10
`11
`12
`13
`14
`15
`16
`17
`18
`19
`20
`21
`22
`23
`24
`25
`26
`27
`
`IPR2023-00037
`Patent 10,971,140 B2
`more. So you could actually have a situation or simply two processors
`accessing a memory would satisfy the claim.
`
`
`Now it happens to be that Chen has multiple clusters of
`processors with a plurality of processors in it, so we did not feel the need to
`have that discussion in the petition, because Chen teaches numerous clusters.
` But your question is well taken.
`
`
`JUDGE SMITH: Well, I'm bringing that out because I guess I
`want to know a couple of things.
`
`
`MS. BAILEY: Okay.
`
`
`JUDGE SMITH: I guess the first thing is did the petition make
`that argument that one cluster with two processors would be enough? So
`that's my first question.
`
`
`MS. BAILEY: The petition did not --
`
`
`JUDGE SMITH: Did the petition make that argument?
`
`
`MS. BAILEY: So sorry. I did not mean to talk over you, Your
`Honor.
`The petition did not make that argument because Chen already
`
`
`teaches multiple clusters of processors with a plurality of processors within
`each cluster.
`
`
`JUDGE SMITH: Chen already teaches it. I see.
`
`
`MS. BAILEY: Yeah.
`
`
`JUDGE SMITH: Yeah, I bring that up because on Page 14 of
`the petition, it looks like it's citation to the declaration Paragraph 72, where
`your expert cites to it looks like six different references that talk about this
`cluster processing with one cluster.
`
`
`Even Jiang by itself, if you look at, for example, Column six of
`Jiang talks about -- this is around Line 41. The CPU may include one or
`
`
`
`
`11
`
`
`
`

`

`1
`2
`3
`4
`5
`6
`7
`8
`9
`10
`11
`12
`13
`14
`15
`16
`17
`18
`19
`20
`21
`22
`23
`24
`25
`26
`27
`
`IPR2023-00037
`Patent 10,971,140 B2
`more processors. And then along with the testimony of your expert from
`Paragraph 72, seems to make it explicit what's already implicit, that the
`cluster of two processors is already if not explicitly disclosed, at least taught
`by Jiang in a knowledge of a person of ordinary skill.
`
`
`And so it wouldn't be necessary to reach all of the arguments
`against the combination, including Chen raised by Petitioner. But it sounds
`like you're saying that it's not possible for us to support a decision based on
`Jiang in view of the knowledge of --
`
`
`MS. BAILEY: So --
`
`
`JUDGE SMITH: -- a person of ordinary skill.
`
`
`MS. BAILEY: Again, I apologize for speaking over. I thought
`you were done.
`
`
`We do note that Jiang teaches a multiprocessor situation -- or
`architecture in our petition. We also realize that Chen has the clusters of
`processors, and so we mapped to Chen. But we also did map that Jiang
`teaches a multiprocessor architecture, as you well pointed out as discussed in
`the declaration.
`
`
`JUDGE SMITH: But I guess to my question, to the extent that
`your primary reference by itself either anticipates or in combination with the
`knowledge of a person of ordinary skill renders this obvious, can we use that
`as a basis for our decision, or do we have to reach the combination of Jiang
`and Chen?
`MS. BAILEY: The petition does talk about how -- and I'm
`
`
`looking at Page 17 through 18, referring again to Jiang's multiprocessor
`system. And it cites the portions that you've noted in addition to a couple of
`additional citations to Jiang and discusses that Jiang alone does teach the
`multiprocessor situation.
`
`
`
`
`
`12
`
`
`
`

`

`1
`2
`3
`4
`5
`6
`7
`8
`9
`10
`11
`12
`13
`14
`15
`16
`17
`18
`19
`20
`21
`22
`23
`24
`25
`26
`27
`
`IPR2023-00037
`Patent 10,971,140 B2
`
`
`JUDGE SMITH: Yeah, but I guess I'm asking you is that
`within the scope of the ground presented in a petition, or would it be beyond
`the ground that you presented for?
`
`
`MS. BAILEY: The ground presents the combination of Jiang
`and Chen.
`JUDGE SMITH: Okay. Thank you.
`
`
`MS. BAILEY: I think that answers your question, Your Honor.
`
`
`JUDGE SMITH: Yes. Thank you.
`
`
`MS. BAILEY: Okay.
`
`
`JUDGE TURNER: Before you go on, this is Judge Turner.
`
`
`Just a quick question. I think if I understand the Petitioner's position, Jiang
`does the pruning basically when it gets done with the tree search engine
`functionality, right?
`
`
`MS. BAILEY: Yes.
`
`
`JUDGE TURNER: But the claim requires, and it's not a
`method claim, but the claim requires that the overall speech recognition
`system circuit is what actually provides the initial score, right?
`
`
`MS. BAILEY: Correct. But I want to make sure that I am
`answering your particular --
`
`
`JUDGE TURNER: I don't think I've asked a question yet.
`
`
`MS. BAILEY: Okay.
`
`
`JUDGE TURNER: So I'm laying out a predicate here, so
`maybe I'm perhaps missing something.
`
`
`But what you're relying on for 1(e) is the pruning, so that to
`continue processing, if you're doing that individually in all these different
`portions in Chen, as you sort of diagrammed out, do you actually get an
`initial score, or do you have just sort of an interim score, then you do the
`
`
`
`
`13
`
`
`
`

`

`IPR2023-00037
`Patent 10,971,140 B2
`pruning, and then you come back with an initial score, and then you
`understand my question, whether what you've asserted in terms of the
`combination, does it meet the claim?
`
`
`MS. BAILEY: I do, and it does. And I want to make sure that
`you understand --
`
`
`JUDGE TURNER: Sure.
`
`
`MS. BAILEY: -- Jiang and then relative to the claim. So
`Jiang, at Column 8, then also talks a bit in at the top of Column 11, the tree
`search algorithm, when you're doing a tree search, the initial score is
`generated for that particular phoneme probability, and the final score is
`generated on the tree. And the final score is that initial score in combination
`with all the scores previous because this is part of the word recognition.
`
`
`After the scores are generated, then Jiang teaches the pruning is
`done. You prune after you have a final score determination for that
`particular phoneme in Jiang and in any tree search algorithm. So the initial
`score is determined, then the final score is determined. And then based on
`whether that final score is above or below the pruning threshold, the
`remainder of the branch will be pruned or not.
`
`
`In the petition's mapping, that tree search algorithms replicated
`across all the processors. The claim ends with the final score determination.
` The claim doesn't say that pruning is then done. In other words, pruning is
`not a part a limitation on this claim.
`
`
`So what happens in Jiang after the final score determination, the
`pruning, because it's not claimed, it's not discussed for purposes of the
`combination. Again, the claim ends with that final score determination that's
`taught at Jiang at Column 8, the lower half. What happens after that is not
`
`1
`2
`3
`4
`5
`6
`7
`8
`9
`10
`11
`12
`13
`14
`15
`16
`17
`18
`19
`20
`21
`22
`23
`24
`25
`26
`
`
`
`
`
`14
`
`
`
`

`

`IPR2023-00037
`Patent 10,971,140 B2
`claimed. Now Jiang does go on and describe what happens after that, the
`pruning that's performed.
`
`
`Does that answer your question, Your Honor?
`
`
`JUDGE TURNER: I think it does, but that does lead to a
`second question. So does the overall claim, the speech recognition circuit,
`does it actually determine, quote, unquote, an initial score, or is it
`determining multiple initial scores and multiple final scores? I guess that's
`my question.
`
`
`MS. BAILEY: So in speech recognition, there's going to be
`thousands of trees --
`
`
`JUDGE TURNER: Sure.
`
`
`MS. BAILEY: -- being run, and each of those trees will have
`an initial and final score determined for the utterance that just came through.
` So 10-millisecond frame is going to do initial and final score. At that point,
`the trees are constantly changing as the utterances come in.
`
`
`So after the final score is calculated, then the tree search
`algorithm will prune, if pruning is performed. Pruning doesn't always get
`performed. It will then prune, and then the next utterances come in and it
`will look at the next node on the tree. This claim ends with that initial and
`final score being determined. The claim doesn't actually talk about how you
`determine the word. It doesn't recite that.
`
`
`JUDGE TURNER: Okay.
`
`
`MS. BAILEY: Am I answering your question, Your Honor?
`
`
`JUDGE TURNER: No, I think you are. Yes.
`
`
`MS. BAILEY: Okay. Moving back to DX-8, I want to quickly
`address some of Zentian's arguments on memory access.
`
`1
`2
`3
`4
`5
`6
`7
`8
`9
`10
`11
`12
`13
`14
`15
`16
`17
`18
`19
`20
`21
`22
`23
`24
`25
`26
`
`
`
`
`
`15
`
`
`
`

`

`IPR2023-00037
`Patent 10,971,140 B2
`
`
`Zentian argues that the memory access would require extensive
`communication. Now Zentian doesn't describe what it means by extensive
`communication, but a takeaway that I want to make sure that the board
`understands is that there does not need to be communication of a cluster's
`processors to another cluster's processors in order to do the initial and final
`score determination.
`
`
`Each processor actually operates independently of itself
`because it's performing the entire tree search algorithm to determine the
`initial and final score. So the processors don't all have to communicate with
`each other, and nothing in Chen actually precludes the processors from
`communicating with each other. It's just simply not required in the claim.
`
`
`JUDGE SMITH: So let me just follow up on that point. So
`that claim, the initial score is used to determine whether to continue
`processing?
`MS. BAILEY: Yes.
`
`
`JUDGE SMITH: You're saying that the circuit doesn't have to
`
`
`do that? There's just something somewhere that uses the initial score?
`
`
`MS. BAILEY: No. The speech recognition --
`
`
`JUDGE SMITH: It --
`
`
`MS. BAILEY: Oh, I'm so sorry again. The speech recognition
`score does determine the initial and final score determination. What I'm
`saying is that each processor running Jiang's tree search algorithm also
`determines the initial and final score.
`
`
`JUDGE SMITH: Okay. I see what you're saying. So the fact
`that it says it's used to determine -- it's implicit in there that it is -- by the
`speech recognition circuit is implicit in that Claim 1(e)?
`
`1
`2
`3
`4
`5
`6
`7
`8
`9
`10
`11
`12
`13
`14
`15
`16
`17
`18
`19
`20
`21
`22
`23
`24
`25
`26
`
`
`
`
`
`16
`
`
`
`

`

`1
`2
`3
`4
`5
`6
`7
`8
`9
`10
`11
`12
`13
`14
`15
`16
`17
`18
`19
`20
`21
`22
`23
`24
`25
`26
`27
`
`IPR2023-00037
`Patent 10,971,140 B2
`
`
`MS. BAILEY: The initial score is used to determine the final
`score. Recall where I just discussed in Jiang, and let's take it back to Jiang.
`Jiang discusses determining a score of the phoneme that is most likely, and
`that's based on just the utterance that it's hearing, but then it determines a
`final score by looking at the initial score along with all of the scores that are
`further up on the branch, because now it's doing word recognition.
`
`
`And so it's saying, okay, what is the lexicon? What is a likely
`word, based on the three phonemes that we just recognized? Is this a likely
`phoneme based on the initial score? So Jiang already teaches the initial and
`final score determination.
`
`
`JUDGE SMITH: Okay.
`
`
`MS. BAILEY: Okay. With respect to Zentian's task, let's move
`to DX-9. Zentian has a series of tasks that would have to be performed in
`the combination. I note that task-sharing, resolving bottlenecks, and
`messaging strategy were just identified but not otherwise discussed by
`Zentian.
`And then I also note that a POSITA would know how to
`
`
`address these issues due to, again, the overwhelming prior art of record
`showing that POSITAs at the time were doing speech recognition with
`parallel processing.
`
`
`JUDGE SMITH: It seems like a lot of the arguments on this
`point are the same as the arguments in the previous case; is that right?
`
`
`MS. BAILEY: You are correct, Your Honor. They are very
`similar arguments and I do feel like I may be repetitive. Actually, I will
`reserve my remaining time for rebuttal if there are no further questions.
`Thank you.
`
`
`
`JUDGE TURNER: Thank you.
`
`
`
`
`17
`
`
`
`

`

`1
`2
`3
`4
`5
`6
`7
`8
`9
`10
`11
`12
`13
`14
`15
`16
`17
`18
`19
`20
`21
`22
`23
`24
`25
`26
`27
`
`IPR2023-00037
`Patent 10,971,140 B2
`
`
`JUDGE OGDEN: Yes. Thank you, Counsel.
`
`
`We'll move now to Patent Owner. Mr. Noroozi, would you like
`to reserve surrebuttal time?
`
`
`MR. NOROOZI: Yes, Your Honor. 15 minutes, please?
`
`
`JUDGE OGDEN: Okay. There's 30 minutes on the timer and
`you can begin when you're ready.
`
`
`MR. NOROOZI: Thank you, Your Honors, and may it please
`the board.
`The board has, I think, struck on the key issues with respect to
`
`
`the arguments on Limitations 1(d) and 1(e). However, I think there's a
`significant misunderstanding on the correct claim language interpretation
`that has arisen from Petitioner's arguments, and I want to clarify that. And I
`think, once we see what the claims require, then everything comes together.
`
`
`So I'd like to start with the language of the claims themselves.
`And Judge Turner, you touched on this question, and I know that it's one
`that's understood by all the judges in terms of its relevance. The Claim 1(d)
`and then 1(e), which follows it, is directed to performance by the speech
`recognition circuit and is not directed to what is done on any individual
`processor within the clusters. And this is really important to understand,
`because what happens in the claim and in the teachings of the patent with
`respect to the initial score determination happens at two levels.
`
`
`There is activity at the level of the individual processors, where
`aspects of what becomes the initial score are calculated, and then that --
`those calculations from the individual processors are aggregated in the
`results memory and used by the search controller to actually calculate what
`becomes the initial score across all of the processors at the search controller
`with respect to Limitation 1(d).
`
`
`
`
`
`18
`
`
`
`

`

`1
`2
`3
`4
`5
`6
`7
`8
`9
`10
`11
`12
`13
`14
`15
`16
`17
`18
`19
`20
`21
`22
`23
`24
`25
`26
`27
`
`IPR2023-00037
`Patent 10,971,140 B2
`
`
`And I know I'm making assurances, but I'm going to show them
`to you in the patent. And that's also where, with respect to Limitation 1(e),
`the determination whether to continue processing or not is also made. So, if
`we look at Column 10 of the 140 Patent --
`
`
`Yes, that would be helpful, if we could publish that.
`
`
`Column 10 of the 140 Patent talks about the operation of the
`search controller with respect to the flow diagram of Figure 6 of the patent.
`And what Column 10 explains is that the search controller assigns out the
`relevant portions of the tree to each individual processor. Each individual
`processor comes up with its own temporary lexical tree score. These
`temporary lexical tree scores are then fed from each individual processor not
`to the cluster memories but to one different common memory, which is the
`results memory.
`
`
`From this results memory, the search controller obtains the
`different temporary lexical tree scores that came from each individual
`processor, and it uses that to form the overall initial score or the overall
`temporary lexical tree score for that entire level of recognition that was
`being performed within that instance. And we can see this in the difference
`between Figures 4 and Figure 6. Let me go to Figure 2 for a minute just so
`we can see it at a very high level.
`
`
`So if we could rotate Figure 2. If you right click, I think it'll
`give you a rotate option. No. But I think the board has Figure 2 available to
`it.
`You'll see that there's a Search Controller 27.
`
`
`And we can just keep it as it is.
`
`
`There's a Search Controller 27, and it shows, with this up and
`
`
`down arrow, that the search controller is in communication with the
`
`
`
`
`19
`
`
`
`

`

`1
`2
`3
`4
`5
`6
`7
`8
`9
`10
`11
`12
`13
`14
`15
`16
`17
`18
`19
`20
`21
`22
`23
`24
`25
`26
`27
`
`IPR2023-00037
`Patent 10,971,140 B2
`processors. The instructions are sent to each processor of what to score, then
`the scores out of the processors back up to the results memory, which I think
`is number 25. And then you can see an arrow going back from the results
`memory to the search controller.
`
`
`That's the high level of the process I was describing, and it's at
`that point that the initial score of Limitation 1(d) occurs and the
`determination whether to proceed to either prune or to continue scoring.
`That's what Limitation 1(e) is talking about. That step of Limitation 1(e)
`also happens at the Search Controller 27. Consistent with the claim, which
`says that it's the speech recognition circuit that does Steps 1(d) and 1(e) and
`not each one individual processor itself.
`
`
`Now if we could go to Figure 4.
`
`
`JUDGE OGDEN: Counsel, doesn't the speech recognition
`circuit include all of the parts, including the other processors?
`
`
`MR. NOROOZI: That's right, Your Honor. And that's why I
`think it's important to look at Figure 4 and 6, because I think what you're
`touching on is, well, it says speech recognition circuit. That could be
`something more than any one individual processor, but it could also just
`comprise the work of one individual processor, so how do we know it has to
`be one or the other, or could it be both? Am I understanding the gist of your
`question, Your Honor?
`
`
`JUDGE OGDEN: Yeah. Go ahead.
`
`
`MR. NOROOZI: Yes. So let's take a look at Figure 4. Figure
`4 shows the flow diagram of the temporary score determination in one
`processor. And if we look at the flow in Figure 4, we see that, at the end of
`it, there is no pruning determination or not.