Trials@uspto.gov
`571-272-7822
`
`
`
`
`Paper No. 33
`Entered: April 11, 2024
`
`UNITED STATES PATENT AND TRADEMARK OFFICE
`________________
`
`BEFORE THE PATENT TRIAL AND APPEAL BOARD
`________________
`
`APPLE INC.,
`Petitioner,
`
`v.
`
`ZENTIAN LIMITED,
`Patent Owner.
`________________
`
`IPR2023-00035
`Patent 10,062,377 B2
`________________
`
`Record of Oral Hearing
`Held: March 12, 2024
`________________
`
`
`
`
`Before KEVIN F. TURNER, JEFFREY S. SMITH, and
`CHRISTOPHER L. OGDEN, Administrative Patent Judges.
`
`
`
`
`
`
`
`
`
`
`
`
`571-272-7822
`
`
`
`
`Paper No. 33
`Entered: April 11, 2024
`
`UNITED STATES PATENT AND TRADEMARK OFFICE
`________________
`
`BEFORE THE PATENT TRIAL AND APPEAL BOARD
`________________
`
`APPLE INC.,
`Petitioner,
`
`v.
`
`ZENTIAN LIMITED,
`Patent Owner.
`________________
`
`IPR2023-00035
`Patent 10,062,377 B2
`________________
`
`Record of Oral Hearing
`Held: March 12, 2024
`________________
`
`
`
`
`Before KEVIN F. TURNER, JEFFREY S. SMITH, and
`CHRISTOPHER L. OGDEN, Administrative Patent Judges.
`
`
`
`
`
`
`
`
`
`
`
`
`
`IPR2023-00035
`Patent 10,062,377 B2
`
`
`
`APPEARANCES:
`
`ON BEHALF OF THE PETITIONER:
`
`
`CRISTINA CANINO, ESQ.
`Erise IP
`Cristina.canino@eriseip.com
`5299 DTC Blvd, Ste. 1340
`Greenwood Village, CO 80111
`(913) 777-5600
`
`JENNIFER BAILEY, ESQ.
`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, ESQ.
`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 1:52 p.m., via video teleconference.
`
`
`
`
`
`
`2
`
`
`
`IPR2023-00035
`Patent 10,062,377 B2
`
`
`P R O C E E D I N G S
`- - - - -
`JUDGE TURNER: We’re going to go on the record. Thank
`
`you. Okay. Good morning again. This is an oral hearing for IPR 2023-
`0035, involving U.S. Patent 10,062,377. I am Judge Turner, joined by
`Judges Ogden and Smith. All the provisos that I provided in the earlier
`hearing still apply to this hearing. In this hearing, we have a LEAP
`participant who will be delivering arguments on behalf of Petitioner, such
`that Petitioner will have a total of 60 minutes to present its arguments, and
`Patent Owner will have 45 minutes to present its arguments.
` Petitioner will go first and present its case. Thereafter, Patent
`Owner arguments in opposition to the petitioner’s case. If there is any
`rebuttal from Petitioner, we will hear it after Patent Owner’s opposition.
`Finally, we will hear from Patent Owner a surrebuttal if requested.
`Petitioner, when you’re ready, please indicating who is here on Petitioner’s
`behalf, and how much time, if any, you wish to reserve for rebuttal.
` MS. CANINO: Thank you, Your Honors. This is Cristina
`Canino with the law firm of Erise IP on behalf of Petitioner, Apple, Inc.
`With me today is my co-counsel, Jennifer Bailey, and in-house counsel at
`Apple, Jenny Liu. I’d like to take 40 minutes for my main argument, and
`reserve 20 for rebuttal.
`
`JUDGE TURNER: And you may begin when you’re ready.
`Oh, let me -- I’m sorry. We need to also do appearances for Patent Owner.
` MR. NOROOZI: Yes, Your Honors. Kayvan Noroozi from
`Noroozi PC for Patent Owner. With me is Mr. Peter Knops from Noroozi
`PC, as well as Ms. Jessica Bernhardt, from the law firm of Bartlit Beck.
`
`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
`
`3
`
`
`
`IPR2023-00035
`Patent 10,062,377 B2
`
`And I would like to reserve 15 minutes, Your Honors.
`
`JUDGE TURNER: Thank you. And for Amazon?
` MR. CHURNET: Hello, Your Honors. Dargaye Churnet from
`Fenwick & West, representing Amazon.
`
`JUDGE TURNER: Thank you. Ms. Canino, please begin
`whenever you’re ready.
` MS. CANINO: May it please the Board. Thank you, Your
`Honors. I want to start by just thanking you for the opportunity to
`participate in the LEAP program today. As a young practitioner, I’m very
`grateful for these opportunities, so thank you. As we turn to the `377 Patent,
`if there are any particular issues that you have questions on, please let me
`know. I’m happy to jump around and move out of order, to ensure that your
`questions are answered. I do want to note, to the extent Your Honors have
`any questions regarding the pipelining arguments as they pertain to the `377
`Patent, Ms. Bailey will be addressing those. Moving to DX2,
`(INDISCERNIBLE) the petition’s Ground 1 maps Claim 1 with the
`combination of Jiang and Smyth.
` COURT REPORTER: I’m sorry. This is the court reporter.
`I’m having a slight hard time hearing her. She’s kind of faint. I don’t know
`if maybe moving closer -- thank you. That works.
` MS. CANINO: Yes.
`
`JUDGE TURNER: Yes, please. Counsel, please keep your
`mouth close to the microphone.
` MS. CANINO: Okay. I will. I’ll try to speak up. Please let
`me know if you are having trouble hearing me. For now, is this a better
`volume?
`
`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
`
`
`
`IPR2023-00035
`Patent 10,062,377 B2
`
`
` COURT REPORTER: Yes.
` MS. CANINO: Okay.
`
`JUDGE TURNER: Yes. Thank you.
` MS. CANINO: Thank you.
`
`JUDGE TURNER: Can you excuse me for just a second? Mr.
`Noroozi, did you have a point of contention?
` MR. NOROOZI: [pointing out that the timer had not yet been
`activated] --
`
`JUDGE TURNER: I’ll do that in a second, if she goes way
`over. So thank you. Thank you for pointing that out. Please continue, Ms.
`Canino.
`
` MS. CANINO: I’ll try not to go over, Your Honors. Looking
`at DX3, Jiang teaches both feature vectors and optional codewords. More
`specifically, Jiang teaches that the digital audio stream is encoded into a
`feature vector, and this is shown in yellow on DX3. Jiang further teaches
`that for some embodiments, the feature vector is further encoded into a code
`word. This is shown in green on DX3.
` Although Jiang discusses details regarding its distance
`calculations using the phrase code word, Jiang from the very outset
`describes distance calculations more generally, from either a feature vector
`or a codeword, and the petition’s mapping addresses Jiang’s teachings for
`optionally encoding a feature vector, and further explains why a POSITA
`would have understood that Jiang distance calculations from a code word
`were likewise teaching distance calculations from a feature vector.
` Thus, the Board does not need to address the further disputes
`regarding code words or otherwise, because Jiang in itself teaches those
`
`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
`
`5
`
`
`
`IPR2023-00035
`Patent 10,062,377 B2
`
`feature vectors, and optionally code words for some embodiments. Do Your
`Honors have any questions regarding Jiang’s teachings here?
`
`JUDGE TURNER: Sure. Petitioner is not relying on either
`embodiment, or on both embodiments? You’re saying that Jiang teaches
`optionally using either feature vectors or code words. In terms of the
`petition, what is Petitioner relying on? Both, or only one?
` MS. CANINO: Yes. The petition maps to both Jiang’s
`distance calculations from a code word, and Jiang’s distance calculations
`from a feature vector.
`
`JUDGE TURNER: Okay. Thank you.
` MS. CANINO: If there are no further questions, I’ll move to
`DX4. Now, Petitioner’s mapping to Jiang’s distance calculations from a
`code word also satisfy the claims. Zentian’s arguments to the contrary are
`premised on technical misunderstandings regarding code words and feature
`vectors. In order to better inform subsequent discussions, I’d like to take a
`few minutes to just give a high-level overview of how a feature vector is
`encoded through vector quantization into a code word.
` Starting with feature vectors, they are exactly what they sound
`like. They are a vector or multidimensional space, wherein each dimension
`represents a different feature of the uttered speech. These features are also
`referred to as spectral characteristics or cepstral characteristics. A code
`word is the same. It’s a feature vector or a multidimensional space, wherein
`each dimension of that vector represents a different feature of the uttered
`speech, with the only difference being that the representation is more
`efficient, to reduce size of the search space and reduce computation.
` The efficiency that arises from a code word is the result of
`
`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
`
`6
`
`
`
`IPR2023-00035
`Patent 10,062,377 B2
`
`vector quantization. Vector quantization is the process of clustering similar
`feature vectors into a smaller subset of vectors called code words. The
`subset of vectors comprise something that is called a code book. To vector
`quantize a particular feature vector, that input feature vector is compared to
`each code word within the code book to determine its best fit. The output of
`this comparison process is an index pointing to the corresponding code
`word, and this code word represents the original data from the feature
`vector. The index is then used to retrieve the corresponding code word to be
`used in further distance calculations, and in the speech recognition system.
` So to put more succinctly, vector quantization takes a feature
`vector at its input, and outputs another feature vector that is represented
`more efficiently. With that understanding in mind, I will walk through each
`of the reasons why Zentian’s arguments are premised on technical
`misunderstandings.
`
`JUDGE TURNER: This is Judge Turner. So if I have a single
`code word, isn’t that -- I understand you’re saying it doesn’t represent a
`single value, but isn’t it actually a single value computationally?
` MS. CANINO: It is not, Your Honor, and for that I will
`actually direct you to DX7 as a starting point, and I can explain exactly why
`a code word is not a single value. And that’s a big reason why I wanted to
`give the background context for where this concept of a single value comes
`up in Zentian’s arguments, and is, you know, the concept of an index in
`general within vector quantization.
`
`JUDGE TURNER: So give me an example. What’s a code
`
`word?
`
` MS. CANINO: A code word is the vector output from
`
`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
`
`7
`
`
`
`IPR2023-00035
`Patent 10,062,377 B2
`
`(CROSSTALK) --
`
`JUDGE TURNER: I understand, but exemplary. So if I’m,
`you know, considering computationally its invites, how many -- give me an
`example of a code word.
` MS. CANINO: So an example would be if you do have a
`speech utterance, there’s going to be all these different characteristics of that
`speech. For example, pitch. And you can represent that pitch within a
`dimension of a vector. And it can be of any dimension, to your point, and
`the `377 Patent actually talks about this. You can have a feature vector, or in
`our case a code word, be any number of dimensions, from 2 to 39 to even
`greater. So if we have this 39-dimensional vector, each dimension is going
`to represent a different characteristic of that speech, for example pitch. So
`one of those dimensions within the 39-dimensional vector is going to
`represent pitch.
` When you’re looking at a code word -- so let’s say you have a
`39-dimensional feature vector that’s vector quantized. The output of that is
`going to be a 39-dimensional vector, where pitch is still represented as one
`of those dimensions within that vector. That’s what the code word is.
`
`JUDGE TURNER: So can a single code word then be
`equivalent to a particular feature vector?
` MS. CANINO: A single code word is a feature vector, and it
`does -- a single feature vector, at the input of vector quantization, results in a
`single code word output. And that single code word is the same amount of
`dimensions as the original feature vector. You’re not (CROSSTALK) --
`
`JUDGE TURNER: But isn’t the code word a value?
` MS. CANINO: A code word is a vector. I think what --
`
`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
`
`8
`
`
`
`IPR2023-00035
`Patent 10,062,377 B2
`
`
`JUDGE TURNER: (CROSSTALK) -- but isn’t it also a value?
`
`Doesn’t it have a -- I can distinguish this code word from another code word
`because it’s assigned a different value, is it not?
` MS. CANINO: Each dimension has a different value, which
`would make one code word different from another code word. But because
`it is a vector and has multiple dimensions, each of those dimensions will
`have an associated value with it. The code word itself is not a single value.
`It is not a scaler quantity.
`
`JUDGE TURNER: Okay. But then again, I’m still looking for
`-- help me how I’m -- if I’m going to -- I have a code word. Give me a
`representation of a code word, exemplary, because I still think it’s -- it’s
`giving me a value. You’re saying it’s not a value. I understand that you’re
`saying it represents multiple values, and if I’m -- let’s say I have an index. It
`goes to a look up table, and under that look up table, there are all the, you
`know, different attributes there, and I distribute and I say, okay, that’s
`separate from the other indexes that I cite, which is going to give me a
`different value, and it’s going to be appreciable for all of those -- you know,
`for those feature vectors, for all those attributes of the feature vector. But it
`seems like the index is an index. It’s not an indexes. I mean, you can talk
`about code words, but don’t you have a singular code word?
` MS. CANINO: Yes. So I want to clarify, the code word is not
`the index. The code word is the vector that the index points to. The index
`points to a vector, a multidimensional vector, that contains a value at each
`dimension of that vector. And I know you keep asking for an example. I
`think what might be most useful is if I can describe what a code word is
`within the context of these cepstral characteristics. Would that be useful, 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
`
`9
`
`
`
`IPR2023-00035
`Patent 10,062,377 B2
`
`are you looking for --
`
`JUDGE TURNER: Sure.
` MS. CANINO: -- something else? Let’s actually turn to DX8.
`Yes. We’ll start at DX8. So when we’re looking at distance calculations --
`and this is why I think in terms of an example, this is actually going to be
`most informative, when looking at the fact that a code word is not a single
`value. To compute a distance calculation, you have to compare the cepstral
`characteristics or the features of that speech to a vector of the same
`dimensions of that acoustic model. To conduct that comparison, you simply
`cannot be comparing the acoustic model vector to a single value. It would
`not produce a meaningful result, and it would be technically infeasible.
` To conduct the distance calculation to determine the most likely
`phoneme, you have to be comparing a vector of let’s say N size dimensions
`to a vector of the same N size dimensions, in order to obtain the comparison
`that is the distance calculation. If we move back to --
`
`JUDGE TURNER: So just to -- is it fair to say that the code
`word is probably an inept description? Because code word is singular, and
`you’re saying actually a code word is represented by an array or a matrix.
`And so, maybe I guess matrix is still singular, but I refer to a matrix, and it
`contains a multitude of values. But I’m still trying to -- I mean, you’re
`saying that the Patent Owner -- I’ll go back to Slide 7 -- is confusing code
`word index with code word. But they do both contain the word code word,
`so maybe the confusion is understandable?
` MS. CANINO: Yes, and I think that’s why my goal here is to
`add clarification here, because I do understand the confusion. But to your
`question about, you know, well, a code word is singular, in the same way
`
`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
`
`
`
`IPR2023-00035
`Patent 10,062,377 B2
`
`that a feature vector is not plural -- you wouldn’t use feature vectors. You
`would use feature vector as you would code word. Each of those are vectors
`of a multidimensional space that contain a number of values at each
`dimension. Did that help?
`
`JUDGE TURNER: Sure. I guess in terms of going, looking at
`the references, is there something in the reference that you could point out
`that would be helpful, that basically says, look, we’re not talking about a
`code word index. We’re really talking about a code word, and by code
`word, I mean equivalent to a feature vector? Is there something in Jiang, or
`perhaps in Smyth, that says here’s what I mean, and you can’t confuse me
`with an index?
` MS. CANINO: Yes. Absolutely. I think the most useful place
`to look would be Jiang, column 7, lines 30 through 43. This describes
`Jiang’s device receiving code words and comparing them to an acoustic
`model. This section of Jiang is describing Jiang’s distance calculations. As
`I’ve described, to compute that distance calculation, you must be comparing
`two vectors of the same size. So as we read Jiang’s disclosures for a
`distance calculation, it is describing a comparison between code words and
`the acoustic model, which informs us that the code word as taught in Jiang
`for its distance calculations is not an index and is not a single value. It is a
`vector.
`
`JUDGE TURNER: Thank you.
`
` MS. CANINO: Are there any further questions on the issue of,
`I guess indexes or single values, before I move on?
`
`JUDGE SMITH: I do. Just real quickly, you know, in your
`opening Slide 3, you mentioned Jiang teaches feature vectors. It also
`
`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
`
`11
`
`
`
`IPR2023-00035
`Patent 10,062,377 B2
`
`teaches code words, and it uses either one for the distance calculations. I
`think that, you know, the arguments about code words, both from you and
`Patent Owner, are important, to the extent that the petition is not using the
`feature of Jiang to teach the claim limitation. It’s not clear from your slide if
`you’re saying that the petition is -- you know, I think you mentioned earlier,
`you’re saying the petition is relying on both feature vectors and code words?
`If the petition is just relying on feature vectors, then all the arguments
`related to code words, we don’t need to reach those. Is that right?
` MS. CANINO: That is correct.
`
`JUDGE SMITH: And then what -- can you tell me real
`quickly, where does the petition say it could be either the feature vectors or
`the code words of Jiang that map to the claim limitation?
` MS. CANINO: Yes. I have a number of cites for you, and I
`think there’s actually two important categories of citations that I’ll provide
`you. One is where the petition has mapped the optional implementation of
`code words in Jiang, and the second is where the petition has explained why
`a POSITA would have understood Jiang’s distance calculations from a code
`word to likewise teach distance calculations from a (CROSSTALK) --
`
`JUDGE SMITH: I guess that’s not really what I -- I’m kind of
`asking now, where does the petition say you’re using feature vectors, and
`not code words?
` MS. CANINO: Sorry. I have a number of cites here. So
`petition at pages 22 through 25 are going to be most informative, particularly
`on page 25. The petition explains and cites back to Claim 1A the mapping
`of Jiang’s distance calculations to the feature vector.
`
`JUDGE SMITH: But then it goes on to discuss code words. I
`
`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
`
`12
`
`
`
`IPR2023-00035
`Patent 10,062,377 B2
`
`mean, it looks like -- I’m looking at page 25, but yet, you know, I see the
`mapping to code words.
` MS. CANINO: Yes. And the reason for this is because
`throughout the petition mapping -- and I’ll point you to some additional
`citations, particularly in Mr. Schmandt’s declaration, where he explains why
`distance calculations from a code word are the same teachings for distance
`calculations from a feature vector, which is why where you see the phrase
`code word, this arises from the fact that Jiang details its distance calculations
`with the phrase code word. But as I said from the very outset, Jiang
`describes distance calculations from either, and that’s why the petition has
`mapped it in a way that kind of compares and explains at multiple places
`code words and feature vectors.
` The declaration paragraph, Mr. Schmandt’s declaration, Exhibit
`#1003, paragraph 142, maps to Jiang’s distance calculations from a code
`word/feature vector. Likewise, Exhibit #1003, paragraph 146 maps a
`distance calculation comparing a feature vector/code word when discussing
`Jiang’s teachings. Again in the declaration, at paragraph 147, feature
`vectors (or code words). Further, in Exhibit #1003, paragraph 212,
`determining the phoneme of the code word (feature vector).
`
`JUDGE SMITH: Okay. Thank you. I understand.
` MS. CANINO: I’d also like to direct Your Honors to page 24
`of the petition, at the bottom, to page 25 of the petition, where a similar
`mapping is described within the petition, along with the citations to the
`declaration. If Your Honors will turn to DX6, one of Zentian’s primary
`technical misunderstandings is about what the code word represents. We’ve
`talked a lot about the fact that a code word is not a single value, but I’d like
`
`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
`
`13
`
`
`
`IPR2023-00035
`Patent 10,062,377 B2
`
`to clarify exactly what is contained within a code word. And Judge Turner, I
`think this will be responsive to some of your earlier questions as well, to the
`extent I didn’t provide a full enough answer then.
` As I’ve already discussed, a code word is itself a feature vector.
`During the vector quantization process, none of the features themselves are
`lost from the input feature vector to the code word. In other words, as I’ve
`said, if the input feature vector is 39 dimensions, the output code word will
`also be 39 dimensions. Zentian does not dispute that a feature vector
`contains a plurality of cepstral characteristics from the digital audio stream.
` Because a code word is simply a vector quantized feature
`vector, it too will represent the same plurality of cepstral characteristics from
`the digital audio stream. If it didn’t, code words would be useless within the
`context of speech recognition, for the reasons I’ve discussed related to
`distance calculations. The entire point of the distance calculation is to
`compare those cepstral characteristics to an acoustic model vector. To do
`this comparison, the code word contains the plurality of cepstral
`characteristics. The vector quantization process does not remove data -- it
`simply clusters it.
` Zentian also argues that because a code word is constructed
`from template vectors or sample data, that it does not satisfy the claims.
`And again, Zentian is technically incorrect. From a starting point, the code
`word is calculated from the feature vector, which is calculated from the
`digital audio stream. And as I’ve indicated, the code word contains that
`plurality of cepstral characteristics from the speech utterance.
` So the fact that the code word may be built from sample data
`does not change the fact that the code word does contain or comprise the
`
`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
`
`
`
`IPR2023-00035
`Patent 10,062,377 B2
`
`plurality of cepstral characteristics from the digital audio stream. Again, this
`is the entire point of vector quantization and the code book. Do Your
`Honors have any questions regarding the fact that a code word contains
`these cepstral characteristics from the digital audio stream?
`
`I’d like to direct Your Honors back a slide, to DX5, and here I
`just want to emphasize that from the very outset, the petition has mapped
`code words as feature vectors, and the mapping is centered on the fact that
`all a code word is -- is simply a vector quantized feature vector. And the
`pertinent portions of this mapping can be found on petition pages 22 through
`25, though you’ll find additional detail throughout the briefing, as I’ve
`alluded to.
` Moving to DX9, with a better understanding of Zentian’s
`technical misunderstandings, it’s easier to understand why a distance
`calculation from a code word is a distance calculation from a feature vector,
`and that is for the reasons that I’ve discussed as it pertains to the distance
`calculation. Again, to compute the distance calculation, you’ll be comparing
`vectors of the same size dimension. Distance calculation from a code word
`does this. The distance calculation also includes a comparison between
`those cepstral features in the acoustic model.
` That’s the entire point of the distance calculation, and this is
`exactly what a distance calculation from a code word does, in addition to
`determining that most likely phoneme, based on a similarity between the
`code word and the acoustic model. If Your Honors can turn to DX10 --
`
`JUDGE TURNER: And just while we’re on DX9 still, just to
`be clear, Petitioner’s position is that Jiang would teach both of those
`columns? So in terms of --
`
`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
`
`
`
`IPR2023-00035
`Patent 10,062,377 B2
`
`
` MS. CANINO: Correct.
`
`JUDGE TURNER: -- doing an acoustic model versus input
`feature vector or versus code word, that Jiang teaches them both, that
`perhaps the code words are more computationally friendly, but Jiang teaches
`both. Is that --
` MS. CANINO: Yes.
`
`JUDGE TURNER: -- the petitioner’s position?
` MS. CANINO: Yes. Absolutely.
`
`JUDGE TURNER: Thank you.
` MS. CANINO: Thank you. Moving to DX10, I do want to
`quickly touch on the lossiness arguments. Ultimately here, when Zentian
`argues that allowing for code words would impermissibly broaden the claim
`language, this is simply imposing an unclaimed degree of accuracy into the
`claims. And the `377 Patent itself, as I’ve alluded to, contemplates for the
`feature vector a number of dimensions within the vector of effectively any
`range, indicating that the Patent itself does not require any specific claimed
`degree of accuracy. And that’s all Zentian is doing here, by arguing that a
`feature vector would not include a code word.
` Turning to DX11, the last thing I wanted to discuss is the
`petition’s alternate mapping, modifying Jiang to include Smyth’s second
`processor for performing distance calculations. Smyth’s distance
`calculations use feature vectors that are not encoded to code words. Now,
`Zentian does not dispute that Smyth teaches the adequate distance
`calculations to satisfy the pertinent limitations. Zentian exclusively argues
`that a motivation to combine was not provided for such combination, and
`this is simply untrue, and not responsive to the petition’s mapping.
`
`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
`
`
`
`IPR2023-00035
`Patent 10,062,377 B2
`
`
`If Your Honors look to paragraph 183 of Mr. Schmandt’s
`
`declaration, Exhibit #1003, there are a number of motivations to combine
`articulated in this paragraph alone, including that a well-known technique of
`calculating distances by comparing a feature vector to the state of an
`acoustic model would improve Jiang’s system. The petition maps these
`combinations and explains these combinations at pages 40 to 41, so I would
`also direct Your Honors there, because a motivation to combine was
`absolutely provided.
`
`JUDGE SMITH: Okay. Can you tell us real quickly what that
`motivation is to combine?
` MS. CANINO: Yes. There’s a number. One is more efficient
`processing. The other is the substitution of a known element. And this is
`actually a very well-known element, the distance calculations from an input
`feature vector for another element, which would be distance calculations
`from (CROSSTALK) --
`
`JUDGE SMITH: For the more efficient processing, can you
`explain how it’s more efficient to use the vectors, as opposed to, I guess the
`code scoring of Jiang?
` MS. CANINO: Yes. Absolutely. So the combination with
`Smyth’s distance calculations includes the combination adding Smyth’s
`second processor, and this motivation and the more efficient processing is
`articulated across paragraphs 170 to 182 of Mr. Schmandt’s declaration. So
`there’s extensive discussion as to why the processing would be more
`efficient, because you are adding a second processor from Smyth.
` Additionally, because vector quantization is computational, if
`you are taking -- if you are removing the vector quantization process, for
`
`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
`
`17
`
`
`
`IPR2023-00035
`Patent 10,062,377 B2
`
`example, and using distance calculations from the feature vector, you’re
`removing the vector quantization process. But the key part here is that the
`motivation does include the addition of Smyth’s second processor, and that’s
`where the greater efficiency comes in. If there are no further questions, I’ll
`reserve my remaining time for rebuttal.
`
`JUDGE TURNER: Just one quick question. From looking at
`the demonstratives here, it looks like Petitioner is choosing to not make any
`arguments with respect to Claim 2, which I’m not sure -- if like Patent
`Owner were not to make any arguments, you’re not going to be able to make
`any in rebuttal, so --
` MS. BAILEY: May I speak to that, Your Honor?
`
`JUDGE TURNER: Sure.
` MS. BAILEY: If you have questions on Claim 2, I will answer
`them. We are not going to make any arguments during the argument today.
`We will rely on our briefing for that, but we still contend that our briefing
`and all of the arguments presented in that.
`
`JUDGE TURNER: Okay. Understood. Thank you. I don’t
`see we have any more questions. Thirty-one. Thirty-one minutes.
` MS. CANINO: Thank you, Your Honors.
`
`JUDGE TURNER: Mr. Noroozi, you can approach when
`you’re ready.
` MR. NOROOZI: Thank you, Your Honors. Kayvan Noroozi
`for Patent Owner. Your Honors, I do want to start with the Claim 2 issue
`very quickly. So the Claim 2 arguments here have been briefed, and they
`are very similar to the argument that you just heard in the `277 proceeding.
`And so, we believe that at this point, the Board is familiar with the issues
`
`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
`
`18
`
`
`
`IPR2023-00035
`Patent 10,062,377 B2
`
`and the disputes and the evidence. The key difference in this proceeding
`with respect to the Claim 2 issue is that the petitioner relies on Nguyen’s
`teachings for the pipelining instead of Brown, but Petitioner treat the two as
`equivalent, and there is no material distinction for purposes of the
`arguments. So unless the Board has particular questions on the Claim 2 and
`beyond arguments, I will move to the Claim 1 argument on code words and
`feature vectors.
`
`JUDGE TURNER: That’s completely up to you, Mr. Noroozi.
`So if you want to direct arguments against Claim 2, you can, or rest on your
`papers and what you’ve previously argued that’s fine either way.
` MR. NOROOZI: Yes, Your Honor. Thank you. So with
`respect to Claim 2, our fundamental points that are demonstrated by the
`evidence and the briefing are that the petitioner relies on the combination of
`a pipelining approach that is fundamentally incompatible with feedback
`based pruning, whereas the combination that they put together includes
`Jiang’s pruning, and that Jiang’s pruning is necessarily feedback based
`pruning, once Jiang is properly interpreted. And that they have some new
`theories that have been discussed in the context of the `277 Patent.
` Those new theories are both untimely, and they also are
`ultimately unsuccessful. The new theory with respect to substituting a one
`frame delay into the pipeline is not taught by any prior art of record. It also
`is refuted by Dr. Anderson’s testimony at his deposition testimony, starting
`at page 34, line 16, to 36, line 1 in terms of its feasibility, because the tree
`would be morphing along, as I have explained in the other proceeding, and
`at this point you would not be able to make use of that fed back information.
` And then with respect to the modification to remove the
`
`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
`
`19
`
`
`
`IPR2023-00035
`Patent 10,062,377 B2
`
`feedback from Jiang’s overall pruning approach, there’s a lot of evidence
`from the prior art, and
Accessing this document will incur an additional charge of $.
After purchase, you can access this document again without charge.
Accept $ Charge
Still Working On It
This document is taking longer than usual to download. This can happen if we need to contact the court directly to obtain the document and their servers are running slowly.
Give it another minute or two to complete, and then try the refresh button.
A few More Minutes ... Still Working
It can take up to 5 minutes for us to download a document if the court servers are running slowly.
Thank you for your continued patience.
This document could not be displayed.
We could not find this document within its docket. Please go back to the docket page and check the link. If that does not work, go back to the docket and refresh it to pull the newest information.
Your account does not support viewing this document.
You need a Paid Account to view this document. Click here to change your account type.
Your account does not support viewing this document.
Set your membership
status to view this document.
With a Docket Alarm membership, you'll
get a whole lot more, including:
- Up-to-date information for this case.
- Email alerts whenever there is an update.
- Full text search for other cases.
- Get email alerts whenever a new case matches your search.
One Moment Please
The filing “” is large (MB) and is being downloaded.
Please refresh this page in a few minutes to see if the filing has been downloaded. The filing will also be emailed to you when the download completes.
Your document is on its way!
If you do not receive the document in five minutes, contact support at support@docketalarm.com.
Sealed Document
We are unable to display this document, it may be under a court ordered seal.
If you have proper credentials to access the file, you may proceed directly to the court's system using your government issued username and password.
Access Government Site
