Trials@uspto.gov
`571-272-7822
`
`
` Paper No. 31
`Entered: April 10, 2024
`
`
`
`UNITED STATES PATENT AND TRADEMARK OFFICE
`
`
`BEFORE THE PATENT TRIAL AND APPEAL BOARD
`
`
`APPLE INC.,
`Petitioner,
`
`v.
`
`ZENTIAN LIMITED,
`Patent Owner
`
`
`IPR2023-00034
`Patent 7,979,277 B2
`
`
`Held: March 12, 2024
`
`
`
`
`
`BEFORE: KEVIN F. TURNER, JEFFREY S. SMITH, and
`CHRISTOPHER L. OGDEN, Administrative Patent Judges.
`
`
`
`
`
`
`

`

`IPR2023-00034
`Patent 7,979,277 B2
`
`
`APPEARANCES:
`
`ON BEHALF OF THE PETITIONER:
`
`
`JENNIFER BAILEY, ESQUIRE
`Erise IP
`7015 College Blvd. , Ste. 700
`Overland Park, KS 66211
`jennifer.bailey@eriseip.com
`(913) 777-5600
`
`
`
`ON BEHALF OF THE PATENT OWNER:
`
`
`KAYVAN NOROOZI, ESQUIRE
`Noroozi PC
`11601 Wilshire Blvd., Ste. 2170
`Los Angeles, CA 90025
`kayvan@noroozipc.com
`(310) 972-7074
`
`
`The above-entitled matter came on for hearing on March 12, 2024,
`commencing at 11:55 a.m., via video teleconference.
`
`
`
`
`
`
`
`
`
`
`
`2
`
`

`

`IPR2023-00034
`Patent 7,979,277 B2
`
`
`
`1
`
`2
`
`3
`
`4
`
`5
`
`6
`
`7
`
`8
`
`9
`
`P R O C E E D I N G S
`
`- - - - -
`
`JUDGE TURNER: Good morning. This is an oral hearing for IPR
`
`2023-00034, involving U.S. Patent 7,979,277. I am Judge Turner, and I’m
`
`joined by Judges Ogden and Smith. For the benefit of the judges and
`
`opposing counsel, as well as the court reporter, please identify yourself when
`
`you begin your argument, and speak clearly into the microphone. Please do
`
`not speak when others, such as the judges, are speaking.
`
`We have the entire record, including demonstratives. When
`
`10
`
`referring to demonstratives, papers, or exhibits, please do so clearly and
`
`11
`
`explicitly by paper or slide number. Please also pause for a second or two
`
`12
`
`after identifying it, to buy us time to find it. This helps in the preparation of
`
`13
`
`an accurate transcript of the hearing. Please bear in mind that the purpose of
`
`14
`
`the hearing is to present your case based on the arguments and evidence of
`
`15
`
`record. You may not introduce new evidence or arguments.
`
`16
`
`I note that there do not appear to be objections to the
`
`17
`
`demonstratives from either side. Please note that we do not permit standing
`
`18
`
`objections to be raised against opposing parties during their arguments. If an
`
`19
`
`advocate believes that new evidence or arguments are to be raised, he or she
`
`20
`
`should make it known to the panel during their principal argument or
`
`21
`
`rebuttal arguments. Each party will have 45 minutes of total argument time.
`
`22
`
`Petitioner and Patent Owner may reserve time for rebuttal. Petitioner will go
`
`23
`
`first and present its case; thereafter, Patent Owner will offer its opposition to
`
`24
`
`the petitioner’s case, and if there is any rebuttal from Petitioner, we will hear
`
`25
`
`it after Patent Owner’s opposition. Finally, we will hear Patent Owner’s
`
`3
`
`

`

`IPR2023-00034
`Patent 7,979,277 B2
`
`surrebuttal, if requested.
`
`I will endeavor to provide each party with a five-minute warning
`
`through the use of the yellow light during opening arguments, and a two-
`
`minute warning during rebuttal or surrebuttal, also through the yellow light.
`
`Once the red light comes on, you should finish up your present argument, or
`
`let the panel know you wish to extend your arguments into prearranged
`
`rebuttal and surrebuttal time, if available.
`
`Please also note that arguments raised during rebuttal or surrebuttal
`
`must be in response to arguments made by the opposing party. Neither
`
`1
`
`2
`
`3
`
`4
`
`5
`
`6
`
`7
`
`8
`
`9
`
`10
`
`period should be used to initiate new arguments. Okay. With that, I’d like
`
`11
`
`to start off with appearances. For this particular IPR, who is appearing on
`
`12
`
`Petitioner’s behalf, please?
`
`13
`
`MS. BAILEY: Good morning, Your Honors. My name is Jennifer
`
`14
`
`Bailey. I’m from the Law Firm of Erise IP. I have here with me my co-
`
`15
`
`counsel, Cristina Canino, and in-house counsel for Petitioner, Apple, Inc.
`
`16
`
`Jenny Liu.
`
`17
`
`JUDGE TURNER: And would you like to reserve for rebuttal, and
`
`18
`
`how much time?
`
`19
`
`MS. BAILEY: Yes, Your Honor. I’d like to reserve 15 minutes
`
`20
`
`for rebuttal.
`
`21
`
`JUDGE TURNER: Okay. And on behalf of Patent Owner,
`
`22
`
`please?
`
`23
`
`MR. NOROOZI: Good morning, Your Honors. Kayvan Noroozi
`
`24
`
`from Noroozi PC for Patent Owner. With me is Mr. Peter Knops, as well as
`
`25
`
`Ms. Jessica Bernhardt, from the Law Firm of Bartlit Beck. And I would like
`
`26
`
`to reserve 15 minutes as well.
`
`4
`
`

`

`IPR2023-00034
`Patent 7,979,277 B2
`
`1
`
`2
`
`3
`
`4
`
`5
`
`6
`
`7
`
`8
`
`9
`
`JUDGE TURNER: When ready, you may begin, please. Pardon
`
`me. I think we do have an appearance from Amazon. I’m sorry. Pardon
`
`me. (CROSSTALK) –
`
`MR. CHURNET: Hello, Your Honor.
`
`JUDGE TURNER: I’m sorry.
`
`MR. CHURNET: Dargaye Churnet, from Fenwick & West, on
`
`behalf of Amazon.
`
`JUDGE TURNER: Thank you very much. Sorry. Ms. Bailey,
`
`when you’re ready.
`
`10
`
`MS. BAILEY: May it please the Board. Thank you, Your Honors.
`
`11
`
`Let’s first turn to DX3 of Petitioner’s demonstratives. In the ’277 Patent
`
`12
`
`IPR, Zentian does not dispute that the limitations are taught. Zentian’s sole
`
`13
`
`argument is a lack of reasonable expectation of success, because feedback-
`
`14
`
`based pruning is allegedly impossible with pipelining.
`
`15
`
`During today when I say feedback, I’m referring to feedback-based
`
`16
`
`pruning, just because that’s a little bit of a mouthful. Pipelining is a required
`
`17
`
`limitation of the claims. This is Limitation 1f. In contrast, feedback is
`
`18
`
`neither claimed nor mapped in the combination. Zentian makes a series of
`
`19
`
`false assumptions to get to the ultimate lack of reasonable expectation of
`
`20
`
`success argument. If any one of these false assumptions fails, Zentian’s
`
`21
`
`entire argument fails. However, the Board does not ultimately need to
`
`22
`
`decide whether Zentian’s false assumptions are correct, because Zentian is
`
`23
`
`impermissibly bodily incorporating feedback into the proposed combination.
`
`24
`
`This alone is sufficient to reject Zentian’s entire argument.
`
`25
`
`So let’s turn to DX4. In the Sur-reply, Zentian takes the position
`
`26
`
`that pruning taught in Jiang is “by definition requires feedback-based
`
`5
`
`

`

`IPR2023-00034
`Patent 7,979,277 B2
`
`1
`
`2
`
`3
`
`4
`
`5
`
`6
`
`7
`
`8
`
`9
`
`pruning.” This is extremely technically incorrect, and so I think it would be
`
`helpful to provide a brief technical summary of feedback relative to pruning,
`
`to kind of set the stage for why Zentian is being -- their arguments are so
`
`technically incorrect.
`
`So as we know from this set of IPRs, there are three stages of
`
`speech recognition: feature vector extraction, distance calculation, and then
`
`the tree search to recognize words. Pruning within speech recognition is a
`
`process of performing fewer operations, at the expense of reduced accuracy
`
`in the speech recognition. If you prune, you potentially have reduced
`
`10
`
`accuracy, but you increase computational resources.
`
`11
`
`There is pruning at the distance calculation stage, and separately,
`
`12
`
`there’s pruning at the search stage. At the distance calculation stage,
`
`13
`
`pruning reduces the number of phonemes each feature vector is compared to.
`
`14
`
`Pruning at the search stage, in contrast, reduces the branches of the trees that
`
`15
`
`are to be searched. And Jiang talks about this at column 8, lines 52 through
`
`16
`
`64, and column 11, lines 8 through 12. In a lexical tree, each branch
`
`17
`
`represents a combination of possible phonemes that collectively form a word
`
`18
`
`in the lexicon. If a word is on a particular branch, and it’s unlikely, based on
`
`19
`
`the combination of phonemes that have been recognized as uttered, then the
`
`20
`
`branch is pruned. This is how the search phase at the tree stage reduces the
`
`21
`
`lexical search phase.
`
`22
`
`In contrast with feedback-based pruning, the results of the pruned
`
`23
`
`branches from the tree search stage are fed back to the distance calculation
`
`24
`
`stage, so that the distance calculator does not perform distance calculations
`
`25
`
`on a phoneme that does not appear on any of the branches as the next
`
`26
`
`potential utterance. So let me describe what this means. Let me just give a
`
`6
`
`

`

`IPR2023-00034
`Patent 7,979,277 B2
`
`1
`
`2
`
`3
`
`4
`
`5
`
`6
`
`7
`
`8
`
`9
`
`brief example. Let’s take the phoneme F. Let’s assume for the words that
`
`the various trees are attempting to recognize for that next utterance, it is
`
`unlikely that the next utterance is going to include the next possible
`
`phoneme as the phoneme F, because the phoneme F does not show up in any
`
`word of all of the branches of all of the lexicon.
`
`If the recognizer sees that there is no branch on any of the trees
`
`that has the phoneme F upcoming, then the recognizer is going to feed back
`
`this information to the distance calculator. The distance calculator now
`
`knows that for the next utterance, the distance calculator does not have to
`
`10
`
`perform a distance calculation for the phoneme F. Now, if the phoneme F
`
`11
`
`still is on some of the branches for the next potential utterance, then the
`
`12
`
`distance calculator isn’t formed that it still does need to perform a distance
`
`13
`
`calculation for the phoneme F, because phoneme F is still active on some of
`
`14
`
`the branches. That is feedback-based pruning.
`
`15
`
`In contrast, pruning at the search stage reduces the branches of the
`
`16
`
`lexical trees to be searched for word recognition. Pruning at the search stage
`
`17
`
`is performed on the branches of the lexical trees because it is unlikely that
`
`18
`
`the collective set of phonemes that have been previously recognized will
`
`19
`
`lead or correspond to a word on the branch. Importantly, pruning at the
`
`20
`
`search stage can be done without feeding back the pruned search results to
`
`21
`
`the distance calculation stage. This is important, so I want to make sure if
`
`22
`
`the Board has any questions on that concept, please let me know.
`
`23
`
`There is independent benefit in pruning at the search stage alone,
`
`24
`
`to reduce the trees to be searched, and we actually know this from a couple
`
`25
`
`of places. One, we know it from Jiang, and the citations I’ve already cited.
`
`26
`
`Also, the ’319 patent that we discussed yesterday, which is Exhibit #1067 in
`
`7
`
`

`

`IPR2023-00034
`Patent 7,979,277 B2
`
`1
`
`2
`
`3
`
`4
`
`5
`
`6
`
`7
`
`8
`
`9
`
`this particular proceeding, Zentian’s own patent describes pruning at the
`
`search stage without feeding back the results, and Zentian does not dispute
`
`this. So we know that pruning at the search stage has independent benefit of
`
`reducing the lexical tree space, separate and apart from pruning at the
`
`distance calculation stage.
`
`So let’s turn to DX5. So this IPR presents a lot of technical
`
`arguments, and it presents several false assumptions by Zentian to make its
`
`arguments work. But the Board does not need to address these issues to
`
`decline Zentian’s argument, and this is because Zentian is impermissibly
`
`10
`
`bodily incorporating feedback into the mapped combination. The claims do
`
`11
`
`not recite feedback. The petition did not map feedback as part of the
`
`12
`
`combination, and even if Jiang teaches feedback -- which Petitioner
`
`13
`
`disputes, but even if Jiang teaches feedback, it would be a fundamental legal
`
`14
`
`error to import feedback into the combination. Because Jiang’s pruning at
`
`15
`
`the search stage can be performed independently of the feedback, there is no
`
`16
`
`technical reason to make feedback part of the combination. Again, feedback
`
`17
`
`is not claimed, nor is it mapped.
`
`18
`
`I want to bring the Board’s attention to a case, Axonics v.
`
`19
`
`Medtronic, this is 73 F 4th 950, where it is a similar set of facts, and this is
`
`20
`
`responsive to Zentian’s Sur- reply arguments on bodily incorporation. In
`
`21
`
`Axonics, it was regarding a medical device case. The Federal Circuit
`
`22
`
`reversed the Board’s finding of no unpatentability. In Axonics, the primary
`
`23
`
`reference in the obviousness combination was directed to a specific context.
`
`24
`
`It was regarding a nerve of the body. But this specific context was not
`
`25
`
`claimed in the challenged patent, similar to what we have here, where
`
`26
`
`feedback is not claimed.
`
`8
`
`

`

`IPR2023-00034
`Patent 7,979,277 B2
`
`1
`
`2
`
`3
`
`4
`
`5
`
`6
`
`7
`
`8
`
`9
`
`In reversing the Board, the Federal Circuit said that it was “a
`
`fundamental legal error to confine the motivation to combine inquiry to the
`
`primary reference as context, when that context was not claimed.” In
`
`quoting from the decision, the Federal Circuit said that an inquiry is not
`
`whether a relevant artisan would combine a first references feature with a
`
`second references feature to meet requirements of the first reference that are
`
`not requirements of the claims at issue. That’s what’s going on here with
`
`Zentian’s arguments, Your Honor. It is legally incorrect for Zentian to
`
`frame the obviousness query as whether feedback with pipelining is
`
`10
`
`possible, when feedback is not even claimed.
`
`11
`
`12
`
`13
`
`JUDGE TURNER: Counsel, this is Judge Turner.
`
`MS. BAILEY: Yes?
`
`JUDGE TURNER: I think yesterday -- I’ll have to cross-reference
`
`14
`
`transcripts, but yesterday we talked about a separate case, about Netflix --
`
`15
`
`16
`
`MS. BAILEY: Yes.
`
`JUDGE TURNER: -- v. DivX. And I guess part of the query that I
`
`17
`
`have here, would it be probable or likely to say that the petitioner really said,
`
`18
`
`okay, feedback -- let me ask a couple questions, just to get a baseline. Does
`
`19
`
`Jiang talk about feedback? Is there feedback in Jiang of any kind, even if
`
`20
`
`it’s not pruning?
`
`21
`
`22
`
`MS. BAILEY: No, there is not.
`
`JUDGE TURNER: So when Patent Owner talks about, you know,
`
`23
`
`Figure 1, where you have -- you know, they’re saying that we have, you
`
`24
`
`know, this information being sent back, that’s not feedback? That’s
`
`25
`
`distinguishable how?
`
`26
`
`MS. BAILEY: Could you tell me what figure you are referring to?
`
`9
`
`

`

`IPR2023-00034
`Patent 7,979,277 B2
`
`1
`
`2
`
`3
`
`4
`
`5
`
`6
`
`7
`
`8
`
`9
`
`JUDGE TURNER: I am looking at Patent Owner’s 14 of the
`
`demonstratives. Maybe I’m mixing up my figures. That’s completely
`
`possible.
`
`MS. BAILEY: Okay.
`
`JUDGE TURNER: It’s Figure 1, anatomy of a speech recognition.
`
`MS. BAILEY: So what Zentian is doing, that’s not from Jiang.
`
`JUDGE TURNER: Okay.
`
`MS. BAILEY: That’s from a prior art reference that does -- or I
`
`should say a background reference. It’s not even using the combination.
`
`10
`
`This is from the Mathew reference, I believe.
`
`11
`
`12
`
`JUDGE TURNER: Okay.
`
`MS. BAILEY: It is not cited as part of the combination. Feedback
`
`13
`
`is disclosed in Mathew, but it’s for the Sphinx recognizer, that Jiang is not
`
`14
`
`the Sphinx recognizer. So what Zentian is doing is it’s using background
`
`15
`
`references to try to inform what Jiang teaches, which is highly improper.
`
`16
`
`There’s no indication that feedback is taught in Jiang. Now, obviously
`
`17
`
`Zentian believes it does, and it has its arguments for it, but it is our very
`
`18
`
`strong position that Jiang does not teach feedback.
`
`19
`
`JUDGE TURNER: Okay. And just to follow up on that, so
`
`20
`
`because Petitioner’s position is that Jiang doesn’t teach feedback at all, you
`
`21
`
`wouldn’t need to carve out that aspect away from the general disclosure?
`
`22
`
`23
`
`24
`
`MS. BAILEY: That’s --
`
`JUDGE TURNER: For example?
`
`MS. BAILEY: That is correct, Your Honor, but may I add on to
`
`25
`
`that? Everything that I just described on the technical basis, remember that I
`
`26
`
`said that it is important to understand that pruning at the search stage can be
`
`10
`
`

`

`IPR2023-00034
`Patent 7,979,277 B2
`
`1
`
`2
`
`3
`
`4
`
`5
`
`6
`
`7
`
`8
`
`9
`
`done with benefit, independently of pruning of the distance calculation
`
`stage. And pruning at the distance calculation stage is feedback, where the
`
`search results from the search stage are fed back to the distance calculation
`
`stage.
`
`Once you understand that pruning can be done separately at the
`
`search stage with independent benefit -- which is what Jiang teaches,
`
`pruning at the search stage -- then there’s no reason, no technical reason to
`
`then try to read into Jiang feedback, when it’s not otherwise taught. In this
`
`technical point, they are distorting and manipulating. Their arguments are
`
`10
`
`very incorrect on this, and they have no rebuttal whatsoever to the technical
`
`11
`
`points that I just told you.
`
`12
`
`13
`
`14
`
`JUDGE TURNER: Thank you.
`
`MS. BAILEY: Okay.
`
`JUDGE OGDEN: This is Judge Ogden. So just so I understand
`
`15
`
`your position, you’re saying that in Jiang, the pruning is accomplished based
`
`16
`
`on scores that were previously calculated before the search stage? So you
`
`17
`
`calculate all the scores for all the parts of the tree, then you do pruning on
`
`18
`
`that basis?
`
`19
`
`MS. BAILEY: Yes, Your Honor. So not to cross-reference
`
`20
`
`discussions from yesterday, but at column 8, the lower half of column 8 of
`
`21
`
`Jiang talks about determining a score, and it actually determines two scores,
`
`22
`
`and we talk about this in our petition. The first score is determining the
`
`23
`
`likelihood of the phoneme. The second score is determining the likelihood
`
`24
`
`of that phoneme actually being in the word, based on the scores of the
`
`25
`
`previous few phonemes above the branch. And so, in a tree search, once it’s
`
`26
`
`determined that that score is below whatever the pruning threshold is, then
`
`11
`
`

`

`IPR2023-00034
`Patent 7,979,277 B2
`
`1
`
`2
`
`3
`
`4
`
`5
`
`6
`
`7
`
`8
`
`9
`
`the remainder of that branch is pruned. Does that answer your question,
`
`Judge Ogden?
`
`JUDGE OGDEN: Yes. Thank you.
`
`MS. BAILEY: Okay. Okay. So if no more questions, turning
`
`back to DX6. I want to quickly hit that another point is that the Board also
`
`does not need to get to whether Zentian’s numerous faults assumptions are
`
`correct, because the art of record also tells us that if feedback is part of Jiang
`
`-- again, we strongly dispute that, but if Jiang does teach feedback, it would
`
`still be obvious to forego feedback in favor of pipelining.
`
`10
`
`I may come back to that argument, but I want to move to a couple
`
`11
`
`of Jiang’s faults assumptions. Let’s move to DX8. So one of the core false
`
`12
`
`assumptions from Zentian is that feedback is impossible with pipelining.
`
`13
`
`Again, their entire argument on the ’277 is a lack of reasonable expectation
`
`14
`
`of success because Jiang allegedly teaches feedback, and feedback would be
`
`15
`
`impossible with pipelining. That is Zentian’s argument.
`
`16
`
`The petitioner reply showed that feedback was possible with
`
`17
`
`pipelining. Quite a bit of the Petitioner reply is devoted to this. I asked Mr.
`
`18
`
`Schmandt, Petitioner’s expert, on this, and he was adamant that feedback is
`
`19
`
`possible with pipelining. So we have spent quite a bit of time explaining
`
`20
`
`why feedback can be done with pipelining. Now, admittedly, there are two
`
`21
`
`tradeoffs with it, and we discuss those tradeoffs in our briefing and in the
`
`22
`
`declaration.
`
`23
`
`The takeaway here, though, is in the Sur- reply, Zentian essentially
`
`24
`
`capitulates that feedback is possible with pipelining. It moves away from
`
`25
`
`that argument, and instead, Zentian raises two new arguments in response.
`
`26
`
`The first argument is that if feedback is done with pipelining, the speech
`
`12
`
`

`

`IPR2023-00034
`Patent 7,979,277 B2
`
`1
`
`2
`
`3
`
`4
`
`5
`
`6
`
`7
`
`8
`
`9
`
`recognition would not be real time, and the second argument is that the
`
`pipelining would not be the type that was taught in Brown, which is what the
`
`combination relies on for pipelining. Zentian is wrong that the speech would
`
`not be real time.
`
`And let’s turn to DX9. You recall that I just mentioned that there
`
`are two tradeoffs if we do feedback with pipelining. Neither one of those
`
`tradeoffs results in the speech recognition not being real time. The speech
`
`recognition and the output rate is the same as if feedback was not used.
`
`In our briefing, we talk about there being a one frame delay, but
`
`10
`
`this is not a one frame delay of the outfit of the speech recognition. Instead,
`
`11
`
`it is a one frame delay between when the results from the tree search are sent
`
`12
`
`back to the distance calculator in a pipeline recognizer recognizing feedback,
`
`13
`
`versus a recognizer that -- or a serial recognizer, one that did not pipeline.
`
`14
`
`So it’s not a delay in the output of the speech recognition. It’s simply a
`
`15
`
`delay in when the distance calculation receives the pruned results. So the
`
`16
`
`speech is still real time.
`
`17
`
`Moving to DX10, Zentian’s second argument is that Brown’s
`
`18
`
`pipelining approach cannot be used with feedback because the operations are
`
`19
`
`not synchronized with the frame generation rate, which is what Brown
`
`20
`
`teaches. Now, that -- let me back up. By the way, that is discussed by
`
`21
`
`Zentian at their Sur-reply at pages 4 and 11. Zentian explains this argument,
`
`22
`
`saying that the search stage for Frame N would not happen at the same time
`
`23
`
`as the distance calculation stage for Frame N plus one. So Frame N is one
`
`24
`
`frame ahead and plus one. The last stage is the tree search stage. And so,
`
`25
`
`Zentian’s argument is that the search stage for Frame N does not happen at
`
`26
`
`the same time as the distance calculation stage for Frame N plus one, so it’s
`
`13
`
`

`

`IPR2023-00034
`Patent 7,979,277 B2
`
`not Brown’s pipelining.
`
`Turning to DX11, once again Zentian is incorrect on the technical
`
`components in this IPR. This diagram is showing a pipeline recognizer
`
`utilizing feedback. It was prepared by Mr. Schmandt, and Zentian does not
`
`dispute this diagram. Does not comment on whether it’s technically correct
`
`or not, has no discussions whatsoever. This diagram, as you can see at time
`
`T plus one, is showing that the search stage for Frame N is performed at the
`
`same time as the classification stage for Frame N plus one. Because again,
`
`this one frame delay that we have talked about is not a delay in the speech
`
`1
`
`2
`
`3
`
`4
`
`5
`
`6
`
`7
`
`8
`
`9
`
`10
`
`recognition being processed; it is simply a delay from the time the search
`
`11
`
`stage pruned results are fed back to the distance calculator, relative to a
`
`12
`
`serial recognizer. Any questions, Your Honors?
`
`13
`
`JUDGE OGDEN: So this figure on DX11 is within the context of
`
`14
`
`-- excuse me -- of recognizing one particular phoneme, and then after it
`
`15
`
`recognizes that, then does it get a reset or something to -- so the pipeline gets
`
`16
`
`a reset, it goes back and starts a new pipeline?
`
`17
`
`MS. BAILEY: It doesn’t start a new pipeline, because the speech
`
`18
`
`utterances are coming in at what is commonly -- a 10-millisecond is
`
`19
`
`common in speech recognition, so each frame of data represents 10
`
`20
`
`milliseconds. So the speech is being uttered, and it’s being processed. So
`
`21
`
`you can see, you can imagine that on each of those times, T is 10
`
`22
`
`milliseconds, T plus one is 10 milliseconds.
`
`23
`
`Pipelining means that you’re doing two stages at the same time, so
`
`24
`
`in a serial recognizer, you would have to search Frame N, and you would not
`
`25
`
`also be doing the classification on Frame N plus one at the same time. In a
`
`26
`
`pipelined or parallel recognizer, you’re doing two steps on two different
`
`14
`
`

`

`IPR2023-00034
`Patent 7,979,277 B2
`
`1
`
`2
`
`3
`
`4
`
`5
`
`6
`
`7
`
`8
`
`9
`
`frames at the same time, which is why you get such an increased efficiency
`
`with parallel processing and speech recognition. Did that answer your
`
`question, Your Honor?
`
`JUDGE OGDEN: I guess maybe my question was not so much
`
`that, you know, it’s a new pipeline. We’re talking about a new tree. I mean,
`
`in the course of the search stage aspect of that, it’s going to do some work on
`
`a particular tree, right? And then send back for the next -- you know, the
`
`next time step. It’s going to send back results. But for one particular
`
`phoneme, we’re working on a particular tree, right? And then I’m just
`
`10
`
`wondering if you can sort of explain how the tree fits in, and maybe that
`
`11
`
`goes more to the next slide, DX12.
`
`12
`
`MS. BAILEY: Yes. So DX12 and 13 -- and I appreciate that there
`
`13
`
`is a lot of text on those, but they do really fully explain the pipelining aspect.
`
`14
`
`Let me try to answer your question by kind of going back to the tree search
`
`15
`
`stage. So at the tree search stage, an utterance is spoken, and it’s determined
`
`16
`
`what phoneme probability that is, and then it’s determined what word
`
`17
`
`likelihood it is, based on the previously spoken phonemes. Because the tree
`
`18
`
`search is all about word recognition, not phoneme, necessarily,
`
`19
`
`identification.
`
`20
`
`So the branches of the trees are constantly changing with each new
`
`21
`
`10-millisecond utterance, because as soon as the recognizer realizes on a
`
`22
`
`particular branch that word is no longer going to be likely, with pruning, it
`
`23
`
`gets rid of that branch. So they’re constantly evolving, and then it becomes
`
`24
`
`even harder when you put in grammar and sentence structure, which is not
`
`25
`
`relevant to here.
`
`26
`
`In feedback, once that branch is pruned, then that result is fed back
`
`15
`
`

`

`IPR2023-00034
`Patent 7,979,277 B2
`
`1
`
`2
`
`3
`
`4
`
`5
`
`6
`
`7
`
`8
`
`9
`
`to the classifier. So let’s look at search Frame N under the search stage. Do
`
`you see the arrow that’s leading back to the classification stage? And it
`
`points to the top of Frame N plus two, which means that the results from
`
`Frame N are going to be considered for classification of Frame N plus two in
`
`a pipelined recognizer utilizing feedback. That’s the one frame delay that
`
`we were discussing. But the output of search Frame N is still at the same
`
`time that it otherwise would have been. The feedback cannot get employed.
`
`I’m going to pause there and see if I have come close, if I’m doing any type
`
`of good job answering your question.
`
`10
`
`JUDGE OGDEN: Yes, I think that’s helpful. I think that helps me
`
`11
`
`understand a little bit better.
`
`12
`
`13
`
`14
`
`MS. BAILEY: Okay.
`
`JUDGE OGDEN: Thank you.
`
`MS. BAILEY: Okay. I’m going to move to DX14. If you have
`
`15
`
`any questions on DX12 and 13, let me know. I actually think because
`
`16
`
`Zentian has essentially admitted in its Sur-reply that feedback is possible
`
`17
`
`with pipelining, the relevance of the work that we did in this Petitioner reply
`
`18
`
`does not seem as appropriate for here. In other words, Zentian capitulates
`
`19
`
`that feedback is possible with pipelining.
`
`20
`
`So turning to DX14, the next false assumption by Zentian that I
`
`21
`
`want to discuss is that the petition’s combination requires Jiang’s pruning.
`
`22
`
`So Jiang teaches that the pruning is optional, and that’s at column 8, lines
`
`23
`
`52. Jiang’s pruning at the search stage was referred to in the petition’s
`
`24
`
`motivation to combine. The petition’s motivation used Jiang’s pruning as
`
`25
`
`the jumping off point for the motivation to combine. Now, obviously the
`
`26
`
`claims don’t recite pruning, so pruning is not mapped as a limitation in the
`
`16
`
`

`

`IPR2023-00034
`Patent 7,979,277 B2
`
`1
`
`2
`
`3
`
`4
`
`5
`
`6
`
`7
`
`8
`
`9
`
`combination, because the claims don’t recite it.
`
`Jiang teaches a desire to increase computational savings, which is
`
`what we want in a speech recognizer. Jiang does that by pruning or
`
`adjusting the threshold for pruning. The problem with pruning, which Jiang
`
`also acknowledges, is that it reduces the accuracy of the speech recognition,
`
`because now you’re not considering as many possible branches for the word
`
`recognition. So when pruning is used, a POSITA would look for another
`
`way to both increase the accuracy -- in other words, don’t do pruning -- but
`
`also increase your computational resources, which is the advantage of
`
`10
`
`pruning.
`
`11
`
`So Jiang’s pruning for increasing computational savings was used
`
`12
`
`as the impetus in paragraph 194 of Schmandt’s declaration, Exhibit #1003.
`
`13
`
`Jiang’s pruning for increasing computational savings was used as the
`
`14
`
`impetus for another way of increasing computational savings, but not having
`
`15
`
`the disadvantages of pruning.
`
`16
`
`Turning to DX15, Zentian’s primary basis for arguing that Jiang’s
`
`17
`
`pruning is part of the combination -- because again, the pruning is optional --
`
`18
`
`is this question and answer from Mr. Schmandt. Mr. Schmandt’s answer to
`
`19
`
`Zentian’s question was simply that pruning is part of the combination, that it
`
`20
`
`was used as part of the motivation to combine, to show that there would be
`
`21
`
`another way -- if you were pruning, when pruning, there would be another
`
`22
`
`way to increase computational savings, by not doing pruning and doing
`
`23
`
`pipelining.
`
`24
`
`If there are no further questions, I’d like to save my remaining time
`
`25
`
`for rebuttal. Thank you.
`
`26
`
`JUDGE TURNER: Sorry about that, Judge Smith? Yes. We are
`
`17
`
`

`

`IPR2023-00034
`Patent 7,979,277 B2
`
`now going to hear from Patent Owner.
`
`MR. NOROOZI: Thank you, Your Honors. Kayvan Noroozi for
`
`Patent Owner. Your Honors, I’d like to clear up quite a bit of potential
`
`confusion that I believe could have been created by Petitioner’s argument
`
`here. I hope that it has not caused confusion to the Board, but I’m confident
`
`that we’ll be able to sort through the facts with actual evidence, actual
`
`statements from the prior art, actual knowledge of how these things work,
`
`and demonstrate to you what’s really going on.
`
`There has been no distortion or manipulation from our side.
`
`1
`
`2
`
`3
`
`4
`
`5
`
`6
`
`7
`
`8
`
`9
`
`10
`
`We’ve tried to engage with the arguments and the combination that the
`
`11
`
`petitioner has put forth, based on actual evidence. Let me start with the
`
`12
`
`assertion that we don’t have any rebuttal to their positions. That’s absolutely
`
`13
`
`incorrect. I think you’ll see that shortly, as well as the assertion that we’ve
`
`14
`
`admitted that feedback is possible with pipelining. We certainly have not
`
`15
`
`done that in the context that is relevant to this proceeding. So we have to
`
`16
`
`evaluate and assess the feasibility of feedback-based pruning with pipelining
`
`17
`
`in the context of the petition’s combination, not with respect to a theory that
`
`18
`
`comes out of Mr. Schmandt’s ruminations today that is not grounded in any
`
`19
`
`prior art.
`
`20
`
`So what we said in response to Mr. Schmandt’s one frame delay
`
`21
`
`reconciliation of feedback-based pruning with pipelining, his attempt at that,
`
`22
`
`is we don’t even need to engage with this fanciful theory, because he
`
`23
`
`admitted in his deposition that there is not one single piece of prior art that
`
`24
`
`teaches that approach. It’s not been demonstrated that this approach that he
`
`25
`
`comes up with was from the time of the invention; instead, it’s something
`
`26
`
`that he came up with himself today, twenty-some years later, with the
`
`18
`
`

`

`IPR2023-00034
`Patent 7,979,277 B2
`
`1
`
`2
`
`3
`
`4
`
`5
`
`6
`
`7
`
`8
`
`9
`
`hindsight of all of our arguments and the claims and everything else. So it’s
`
`absolutely improper. It’s not part of their theory. They don’t have a one
`
`frame delay theory in the petition. It’s not in the prior art.
`
`And we’ll get to all that, but I just want to clear up that confusion
`
`first of all. Because the actual fact is that we demonstrated that Brown’s
`
`pipelining, which is the one that they rely on, is incompatible with feedback-
`
`based pruning, and that this is taught expressly in the prior art. And they did
`
`not respond to that. Instead, what they did is they pivoted away to a
`
`different pipelining, this one frame delay pipelining, which Mr. Schmandt
`
`10
`
`came up with today, which is not in the prior art, and they said, well, we can
`
`11
`
`think of a way today that maybe you could make feedback-based pruning
`
`12
`
`and pipelining happen together. That’s not responsive to the argument we
`
`13
`
`made.
`
`14
`
`Our argument was the pruning of Jiang and the pipelining of
`
`15
`
`Brown that are part of your combination are not compatible, and we proved
`
`16
`
`that, and they don’t d