US 10971140

I'll research this patent across USPTO and litigation sources.

Let me pin down the specific appellate outcome for 2024-2208, which is the appeal from IPR2023-00037 on this patent.

US 10,971,140 B2 — Verification Summary

Caveat on scope of my search: I searched public web sources (Google Patents, USPTO PTAB/PTACTS document mirrors via Docket Alarm, CAFC docket pages, Justia, Fed Circuit Blog) rather than a live USPTO PatentCenter/CAFC PACER query. Details below are grounded in the sources returned; where I could not confirm something, I say so.

Bibliographic data (confirmed)

It is the sixth member of a continuation chain: 10/503,463 (US 7,587,319) → 12/554,607 (US 8,036,890) → 13/253,223 (US 8,768,696) → 14/309,476 (US 9,536,516) → 15/392,396 (US 10,217,460) → 16/266,265 (US 10,971,140). PCT/GB2003/000459; AU2003208392A1; GB2384901B.

Abstract (as granted)

"A speech recognition circuit comprises an input buffer for receiving processed speech parameters. A lexical memory contains lexical data for word recognition. The lexical data comprises a plurality of lexical tree data structures. Each lexical tree data structure comprises a model of words having common prefix components. An initial component of each lexical tree structure is unique. A plurality of lexical tree processors are connected in parallel to the input buffer for processing the speech parameters in parallel to perform parallel lexical tree processing for word recognition by accessing the lexical data in the lexical memory. A results memory is connected to the lexical tree processors for storing processing results from the lexical tree processors and lexical tree identifiers to identify lexical trees to be processed by the lexical tree processors. A controller controls the lexical tree processors to process lexical trees identified in the results memory by performing parallel processing on a plurality of said lexical tree data structures."

(Note: the abstract describes the 2002 disclosure generally; the granted claim 1 is drafted considerably more broadly and does not recite lexical trees, results memory, or a controller — a point that mattered in the IPR.)

Plain-language overview of the claims

Claim 1 (sole independent claim). A speech-recognition circuit comprising one or more clusters of processors. Each cluster contains (a) a plurality of processors and (b) an acoustic model memory storing acoustic model data. Each processor in a cluster is configured to compute a probability using the acoustic model data in that cluster's acoustic model memory. The circuit is further configured to generate an "initial score" for an audio sample, and that initial score is used to decide whether to keep going and determine a "final score" — where the final-score processing uses a larger amount of model data than was used to produce the initial score. In plain terms: a coarse first-pass computation over a small slice of the acoustic model, used as a gate/beam decision on whether to spend resources on a fuller computation. This is a much broader claim than the original 2002 specification's parallel-lexical-tree architecture.

Claim 2 — the computed probability is an input to evaluating a state transition of a model of states.
Claim 3 — that model is a Hidden Markov Model.
Claim 4 — the probability is computed from a Gaussian mixture model and one or more feature vectors.
Claim 5 — adds a buffer storing one or more feature vectors, coupled to at least one processor of the cluster(s).
Claim 6 — adds a search controller coupled to at least one processor, capable of controlling it to initiate speech-recognition processing in a subset of the processors.
Claim 7 — the clusters include a first and second cluster, with a first acoustic model memory and a distinct and separate second acoustic model memory.
Claim 8 — the first cluster's first acoustic model is coupled to all processors of the first cluster.

Litigation and PTAB / CAFC posture for this patent

• District court: Zentian Ltd. v. Apple Inc., No. 6:22-cv-00122 (W.D. Tex., filed 2022-02-02) and No. 6:22-cv-00123; both transferred in June 2023 to N.D. Cal. as No. 3:23-cv-02921-TLT.
• IPR2023-00037 — Apple Inc. v. Zentian Limited (patent 10,971,140; application 16,266,265; Tech Center 2600). Filed 2022-12-02; instituted 2023-06-12; Final Written Decision 2024-06-11 — "Determining No Challenged Claims Unpatentable" for claims 1–8. Panel: APJs Kevin F. Turner, Jeffrey S. Smith (opinion author), Christopher L. Ogden (concurrence).
• IPR2023-01197 ([[Amazon.com Services](/litigations/by-defendant/Amazon.com%20Services) LLC](/litigations/by-plaintiff/Amazon.com%20Services%20LLC)) — institution decision and grant of joinder entered 2024-01-10, joined to the -00037 proceeding.
• CAFC appeal: Apple Inc. v. Zentian Ltd., No. 2024-2208 (Fed. Cir.) — Apple's appeal from the -00037 Final Written Decision; oral argument held/audio posted 2026-06-04. I did not find any disposition (Rule 36 judgment or opinion) for 2024-2208 in my search results — it appears pending as of the material I retrieved. Treat the appeal's outcome as unverified.

Important disambiguation: the Zentian/Apple 2026 Federal Circuit activity that dominates recent news (e.g., the Sept. 25, 2026 nonprecedential decisions and the "feature vector includes codewords" holding) concerns other patents in the same family — U.S. 10,062,377 (the '377 patent) and 10,839,789 (the '789 patent) in appeals 24-2204, 24-2205, 24-2207, not 10,971,140. Do not attribute those holdings to this patent.

Data-quality flag: the Google Patents "Non-Patent Citations" list for this document includes several terahertz-imaging papers (Woodward, Wallace, Cole, Fitzgerald) that are unrelated to speech recognition — these appear to be a citation-record artifact carried across the family, not meaningful prior art for this patent.

Where I am uncertain

1. The outcome of Fed. Cir. 2024-2208 (the appeal of the '140 patent's IPR) — no disposition found in my searches.
2. Whether any 2026 CAFC docket entry beyond the June 4, 2026 oral argument exists for this specific patent number; I found none, and the September 2026 decisions I located are for sibling patents.
3. Exact current USPTO fee/status entries (Google Patents shows "Expired – Fee Related," which I have not independently re-confirmed against PatentCenter).

I'll research litigation involving US patent 10971140. Let me search multiple sources.

Let me gather more detail on outcomes and the additional cases.

Litigation Involving U.S. Patent No. 10,971,140

Patent under analysis: U.S. 10,971,140 B2 ("Speech recognition circuit using parallel processors"), inventor Mark Catchpole, assignee Zentian Ltd., application 16/266,265, filed 2019‑02‑04, granted 2021‑04‑06, priority 2002‑02‑04. Family ID 9930354. Status: Expired – Fee Related (anticipated expiry February 2023 per Zentian's own complaint).

I located litigation in three forums. Note that the '140 patent was asserted by Zentian together with related patents in the same family groups — U.S. 7,587,319 and 7,979,277, 10,062,377, 10,839,789 — so some docket numbers below cover the '140 alongside sibling patents. I have flagged where that is the case.

---

1. District Court — Zentian Ltd. v. Apple Inc.

The asserted patents in this suit were the '319, '277, '377, '789, and '140 patents. Apple's motion to transfer was granted on private‑ and public‑interest factors; the case moved to N.D. Cal. as 3:23-cv-02921-TLT.

---

2. District Court — Zentian Ltd. v. [Amazon.com, Inc.](/litigations/by-plaintiff/Amazon.com%2C%20Inc.)

The same five patents ('319, '277, '377, '789, '140) were asserted against Amazon. Accused products included Echo, Echo Dot, Echo Studio, Echo Show, Echo Auto, Echo Buds, Echo Frames, Fire TV Cube, Fire TV Stick, and certain Fire Tablets.

---

3. PTAB — Apple Inc. v. Zentian Limited, IPR2023-00037

---

4. PTAB — Amazon.com Services LLC et al. v. Zentian Limited, IPR2023-01197

This petition was substantively identical to Apple's and was resolved via a motion to join the Apple proceeding (granted over Zentian's § 315(b) time‑bar and Fintiv arguments). The Board document text references the joined Apple proceeding by number; the proceeding itself is captioned against the '140 patent.

---

5. Federal Circuit — Apple Inc. v. Zentian Limited (appeal of IPR2023-00037 / IPR2023-01197)

⚠️ Caution: Several companion Federal Circuit appeals arising from the same Zentian–Apple dispute involve the sibling patents, not the '140 — e.g., 24-2204 (IPR2023-00033/'319), 24-2205 ('377 feature-vector construction), and 24-2207 (IPR2023-00036/'789). A Federal Circuit decision dated June 8, 2026 affirmed the PTAB in the case captioning "Zentian Ltd. v. Apple Inc., 2024-2204." Do not attribute that June 2026 affirmance to the '140 patent without confirming the 24-2208 docket directly.

---

Summary of litigation touching the '140 patent

Bottom line: The '140 patent was asserted by Zentian against both Apple and Amazon in W.D. Tex. (with the Apple case transferred to N.D. Cal. and stayed), and was challenged at the PTAB by Apple (IPR2023-00037) and Amazon (IPR2023-01197). Zentian successfully defended all claims 1–8 of the '140 patent at the PTAB, and Apple's appeal of that loss is pending at the Federal Circuit (24-2208).

---

Sourcing notes / limitations:

• I relied on the Unified Patents litigation portal entries reflected on the Google Patents page, PTAB institution and final‑written‑decision documents, CourtListener RECAP filings (N.D. Cal. 3:23-cv-02921, docs. 116 and 125), the Amazon invalidity contentions and complaint in 6:22-cv-00123-ADA, and Docket Alarm/Ex Parte PTAB records.
• I did not have direct PACER access in this session. The outcome/current posture of the Amazon district court case (6:22-cv-00123-ADA) and the merits disposition of CAFC 24-2208 could not be verified and are stated as inconclusive. Per your instruction, I did not auto‑correct any case numbers; the Federal Circuit appeal number for the '140 matter is reported as 24-2208, and I have expressly flagged the risk of confusion with the related 24‑2204/2205/2207 appeals covering sibling patents.

I'll research the PTAB proceedings referenced in the patent record, since the Google Patents family data lists two IPR numbers even though the ODP block reports none.

Let me drill into the FWD outcomes and the Federal Circuit appeals.

Proceedings overview

Two AIA trial proceedings naming US 10,971,140 are on the public record — IPR2023-00037 (Apple, Inc. v. Zentian Ltd.) and IPR2023-01197 ([[Amazon.com Services](/litigations/by-defendant/Amazon.com%20Services) LLC](/litigations/by-plaintiff/Amazon.com%20Services%20LLC) / Amazon Web Services, Inc. v. Zentian Ltd.) — and the picture is: 0 claims canceled, 0 claims sustained-by-partial-FWD, 8 of 8 challenged claims (claims 1–8) found not unpatentable; both proceedings were instituted, neither settled, and the '140 patent is one of the few Zentian family members that came out of the June 2024 FWD round intact. The structured "PTAB proceedings on file" block reports the ODP ingest as empty ("The USPTO ODP API returns no AIA trial proceedings for this patent as of the most recent ingest"); that ingest is stale — the Google Patents family/litigation record attached to this patent lists both IPR2023-00037 and IPR2023-01197 as "filed (Final Written Decision)," and the underlying paper filings are public. The bottom line for a defendant: the troll's asserted claims here are hardened, not dead. Unlike the sibling '319, '277 and '377 patents (all challenged claims held unpatentable on 2024-06-11), a Section 103 attack built on the Jiang/Chen combination already failed once at the Board and is now the subject of a pending (argued) Federal Circuit appeal.

---

IPR2023-00037 — Apple Inc. v. Zentian Limited

• Type: Inter Partes Review (35 U.S.C. § 311(a)), all claims 1–8 of U.S. Patent No. 10,971,140 B2.
• Filed: 2022-12-02.
• Status: Verbatim from the structured/PTAB data: "Final Written Decision – Appealed." Plain English: instituted 2023-06-12, FWD 2024-06-11 against Apple, and Apple filed a Notice of Appeal on 2024-08-12; the appeal (No. 2024-2208) was orally argued at the Federal Circuit on 2026-06-04 and, as of the information retrieved, no disposition has been confirmed.
• Judge panel: Kevin F. Turner, Jeffrey S. Smith, and Christopher L. Ogden, Administrative Patent Judges. Opinion for the Board filed by APJ Smith; a separate opinion concurring was filed by APJ Ogden — i.e., a non-unanimous-in-reasoning panel, which is why the decision is sometimes described in the press as a "split" Board decision.
• Petition grounds: Obviousness under 35 U.S.C. § 103(a) against claims 1–8. The record frames the core combination as Jiang in view of Chen (Zentian's POPR and its 2024-03-04 demonstratives both attack "the Petition['s] … Jiang's CPU-based speech recognition system … combin[ed] with Chen's parallel processing system," and identify the dispositive limitations as 1(d) ("generate an initial score for an audio sample") and 1(e) ("the initial score is used to determine whether to continue processing to determine a final score via processing a larger amount of model data than that was processed to generate the initial score")). Apple's expert was Christopher Schmandt (Ex. 1003); Zentian's was David V. Anderson, Ph.D. (Ex. 2020). I could not confirm from the retrieved sources the complete exhibit-by-exhibit reference list beyond Jiang and Chen, so I am not naming secondary references with certainty.
• Institution decision: Instituted 2023-06-12 on all grounds and all of claims 1–8. The Board found a reasonable likelihood under § 314(a) and declined Zentian's Fintiv discretionary-denial argument (Apple had entered a Sotera stipulation).
• Final Written Decision: Entered 2024-06-11, captioned "JUDGMENT — Final Written Decision Determining No Challenged Claims Unpatentable (35 U.S.C. § 318(a))." Quote: "We determine Petitioner has not shown by a preponderance of evidence that claims 1–8 are unpatentable." No independent claim was canceled and no dependent claim was canceled. At least claim 1 — the only independent claim in the '140 patent — was sustained, with dependents 2–8 falling with it. (The FWD PDF is at the Docket Alarm mirror: https://www.docketalarm.com/cases/PTAB/IPR2023-00037/Apple_Inc._v._Zentian_Limited/docs/06-11-2024-Board/Final_Written_Decision__original-34-Judgment_Final_Written_Decision_.pdf ; institution decision: https://www.docketalarm.com/cases/PTAB/IPR2023-00037/Apple_Inc._v._Zentian_Limited/docs/06-12-2023-Board/Institution_Decision__Grant-10-Granting_Institution_of_Inter_Partes_Review_35_USC_%C2%A7_314.pdf ) I have not verified the FWD's exact claim-by-claim table beyond the judgment caption and the "claims 1–8 … not unpatentable" holding, and I am not going to attribute reasoning to the FWD that I could not read.
• Settlement / termination: No settlement. The proceeding ran full course to FWD. (Zentian and Apple were simultaneously in Zentian Ltd. v. Apple Inc., W.D. Tex. 6:22-cv-00122, transferred to N.D. Cal. 3:23-cv-02921, where the district case was stayed on 2023-07-26 pending the IPRs; the parties stipulated on/around 2024-07 to keep the stay in place "for at least three months" to attempt settlement — https://storage.courtlistener.com/recap/gov.uscourts.cand.[414047](/patent/414047)/gov.uscourts.cand.414047.125.0.pdf . Whether a global settlement ultimately occurred is not confirmed in the sources retrieved.)
• Appeal: Yes. Apple filed its Notice of Appeal 2024-08-12; Federal Circuit No. 2024-2208, Apple Inc. v. Zentian Ltd. Oral argument audio was posted 2026-06-04 (https://www.cafc.uscourts.gov/06-04-2026-2024-2208-apple-inc-v-zentian-ltd-audio-uploaded/ ). Disposition not confirmed as of 2026-09-30 — treat the appeal as the live variable. Note the contrast with the sibling appeals: 2024-2204 (Zentian v. Apple, IPR2023-00033/01192) was affirmed under Fed. Cir. R. 36 on 2026-06-08, and 2024-2205 (Zentian v. Apple, IPR2023-00035, '377 patent) was affirmed on 2026-09-25 in a nonprecedential opinion rejecting Zentian's "feature vector" construction argument (https://law.justia.com/cases/federal/appellate-courts/cafc/24-2205/24-2205-2026-09-25.html ). 2024-2208 is Apple-as-appellant, i.e., the reverse posture.
• Defensive value: This is the proceeding that most helps Zentian and most hurts a defendant. The obviousness theory a well-resourced defendant (Apple, with Amazon riding along) actually ran against claim 1 failed on the merits, so a copycat § 103 attack on the '140 patent has to bring genuinely different art or a different technical theory — Jiang/Chen is now a proven loser. But it is not immunity: the FWD is on appeal, so a defendant today should monitor 2024-2208 and can still argue the Board got it wrong.

---

IPR2023-01197 — Amazon.com Services LLC / Amazon Web Services, Inc. v. Zentian Limited

• Type: Inter Partes Review (35 U.S.C. §§ 311–319), claims 1–8 of the '140 patent, plus a Motion for Joinder under 35 U.S.C. § 315(c) / 37 C.F.R. § 42.122.
• Filed: 2023-07-12.
• Status: Verbatim: "Final Written Decision – Appealed"; industry trackers record the outcome as "Patentable." Plain English: instituted and joined into IPR2023-00037 on 2024-01-10; it did not generate a separate merits verdict, and it resolved with the '140 FWD of 2024-06-11 (no claims unpatentable).
• Judge panel: Same three APJs — Kevin F. Turner, Jeffrey S. Smith, Christopher L. Ogden (opinion authored by APJ Smith).
• Petition grounds: Amazon's petition was, by its own words, "intentionally identical in substance to the Apple Petition" — same claims (1–8), same grounds, same primary expert opinion. Amazon's declarant, Dr. Les Atlas, expressly adopted the Schmandt declaration (Ex. 1043 ¶ 26). The Amazon institution record additionally lists § 103(a) art including U.S. Patent No. 6,374,219 B1; U.S. Patent No. 5,428,803; U.S. Patent Appl. Publ. No. 2001/0053974 A1; U.S. Patent No. 5,983,180 ("Robinson"); and U.S. Patent No. 5,036,539 — I note these as appearing in the Amazon record but flag that I could not verify which of them map to the Jiang/Chen grounds versus separate grounds.
• Institution decision: 2024-01-10 — institution granted on all challenged claims and all grounds, and joinder granted. The Board rejected Zentian's argument that Amazon should be required to file a Sotera stipulation like Apple: "we find no unfairness or prejudice in the fact that Apple has submitted a Sotera stipulation but Amazon has not," noting the Board would decide patentability before the parallel district-court trial date, so Fintiv factors 1–5 "weigh heavily against discretionary denial." Critically for strategy, the Board ordered that "the grounds of unpatentability on which the Board instituted trial in IPR2023-00037 are unchanged and remain the only grounds on which trial has been instituted," that filings by the petitioner parties be consolidated, and that no Amazon-only filings are permitted without Board approval (https://www.docketalarm.com/cases/PTAB/IPR2023-01197/Amazon.com_Services_LLC_v._Zentian_Limited/docs/01-10-2024-Board/Institution_Decision_Joined-10-Institution_Decision__Granting_Institution_of_Inter_Partes_Review_35_USC_sec_314_Granting_Motion_for_Joinder_35_USC_sec_315c_37_CFR_sec_42122.pdf ).
• Final Written Decision: No standalone FWD. Because joinder was granted and the grounds froze, Amazon rode the 2024-06-11 FWD in IPR2023-00037 holding claims 1–8 not unpatentable. Do not cite an "IPR2023-01197 FWD" as a cancellation document — there is no such cancellation.
• Settlement / termination: No settlement; terminated with the joined FWD on 2024-06-11. (Amazon's parallel district case, Zentian Ltd. v. Amazon.com, Inc., W.D. Tex. 6:22-cv-00123, terminated 2024-11-20 per the CAFC related-case statement in the '789 appeal record.)
• Appeal: The '140 FWD it joined is on appeal as 2024-2208 (Apple v. Zentian); trackers list IPR2023-01197 itself as "Final Written Decision – Appealed." I could not confirm whether Amazon filed a separate notice of appeal or was carried by Apple's appeal as a joined party — verify on PTAB E2E / the Federal Circuit docket before relying on it.
• Defensive value: Modest and derivative. The practical point is joinder: a defendant that is time-barred under § 315(b) can sometimes still get into an existing instituted IPR via § 315(c), as Amazon did — but only on the frozen instituted grounds. That door, however, requires an existing live trial, and IPR2023-00037 is now at the FWD/appeal stage, so joinder into it is no longer available.

---

Strategic summary

Claim status on US 10,971,140. There is no canceled claim on this patent. Claims 1–8 were each challenged and each survived the only merits adjudication to date (IPR2023-00037 FWD, 2024-06-11, "No Challenged Claims Unpatentable"). Claim 1 is the sole independent claim and it is the one the Board sustained; claims 2–8 are dependents of claims 1–2 and rise or fall with them. District-court-wise, Apple's invalidity contentions asserted '140 claims 1–6, so the asserted set is a subset of what the Board already upheld. In short: 1–8 SUSTAINED (subject to appeal); none CANCELED; none UNTESTED at the PTAB — though no institution-stage or CBM/PGR challenge has ever been filed, and no § 112 challenge has been adjudicated at the Board (Apple raised written-description/priority attacks only in district court).

Estoppel landscape. § 315(e)(2) estoppel attaches to Apple (and to Amazon, as a joined petitioner who argued the same grounds) upon the FWD — it bars them and their privies, in the district case and before the Office, from raising any ground they raised or reasonably could have raised. Practically: Apple and Amazon cannot re-run Jiang/Chen, nor any obviousness combination of printed art they could have assembled against claims 1–8 in the W.D. Tex./N.D. Cal. cases. That estoppel does not run to unrelated defendants. A new defendant (Samsung, Google, Meta, a chip vendor, etc.) is free to petition against claims 1–8 using fresh art — and should, since the patent has now shown it can beat a sophisticated challenger. Live non-IPR levers for a non-petitioner: (i) § 112 / priority — Apple's contentions expressly attacked entitlement to the 2002-02-04 priority date on written-description and enablement grounds; if that date breaks, pre-2002 art opens up; (ii) claim construction — the family's recent appellate loss (2024-2205, '377, 2026-09-25) turned on claim-construction scope for "feature vector," a term that also appears in the '140 claims (see claim 4 and claim 5), so construction arguments still have traction; (iii) eligibility/§ 101 and damages/expiry arguments, which the Board never reached.

Pattern signals. This is a multi-front, co-ordinated defense against a single patent owner, not a defense-aggregator campaign. Apple filed five IPRs (IPR2023-00033 through -00036 and -00037) covering five Zentian patents ('319, '277, '377, '789, '140); Amazon filed flanking petitions (including IPR2023-01192 and -01197) that were routed into Apple's instituted trials via § 315(c) joinder rather than fought on their own. The results were mixed: Apple won all-claims cancellation on the '319 (IPR2023-00033), '277 (IPR2023-00034) and '377 (IPR2023-00035) patents, and lost on the merits on the '789 (partially) and '140 (entirely). Zentian has been aggressive on appeal, appealing every loss (2024-2204, 2024-2205, and the '277 appeal) and losing — 2024-2204 affirmed under Rule 36 on 2026-06-08 and 2024-2205 affirmed on 2026-09-25; Apple cross-appealed its two losses (the '789 in 2024-2207 and the '140 in 2024-2208). To answer the aggregator question directly: Unified Patents is not a petitioner here. The "Unified Patents" strings on this patent's Google Patents page are the data-source attribution for the litigation/proceeding listings ("Unified Patents Litigation Data" / "Unified Patents PTAB Data"), not a party — do not let the platform's boilerplate mislead you into drafting around a Unified IPR that does not exist.

One more defensive data point worth verifying immediately. The Google Patents record for this patent states an "Anticipated expiration" of 2023-02-04 (20 years from the 2003-02-04 PCT/US filing of the original application; the '140 is a continuation of a § 371 national-stage case) and a current legal status of "Expired - Fee Related." If that term/status is accurate, prospective injunctive relief is off the table and damages are limited to pre-expiry infringement within the § 286 six-year lookback — a point to confirm against the USPTO Patent Center maintenance-fee and term-adjustment record before it is relied on, since I am reading this off a third-party patent database rather than the Office's own term calculation.

Recommended next steps

• Pull and read the FWD before anything else. The controlling document is the 2024-06-11 "Final Written Decision Determining No Challenged Claims Unpatentable" in IPR2023-00037: https://www.docketalarm.com/cases/PTAB/IPR2023-00037/Apple_Inc._v._Zentian_Limited/docs/06-11-2024-Board/Final_Written_Decision__original-34-Judgment_Final_Written_Decision_.pdf . Quote it accurately: the Board held "Petitioner has not shown by a preponderance of evidence that claims 1–8 are unpatentable." That is the sentence to put in front of your client and, if useful, in a Rule 408 / § 285 framing letter — Apple and Amazon brought a full-court press and no claim was canceled.
• Docket-check Federal Circuit No. 2024-2208 (Apple Inc. v. Zentian Ltd.) today. Oral argument was heard 2026-06-04; there is a real possibility of a decision by now that I could not confirm. If the FWD is affirmed, the '140 patent is effectively hardened against § 103 and the case should pivot to § 101, § 112, damages/expiry and non-infringement. If remanded, the Board's obviousness analysis on limitations 1(d)/1(e) reopens and a follow-on petition becomes viable. Docket/audio landing page: https://www.cafc.uscourts.gov/06-04-2026-2024-2208-apple-inc-v-zentian-ltd-audio-uploaded/ ; PTAB papers via USPTO PTAB E2E at https://e2e.uspto.gov/ and https://ptacts.uspto.gov/ .
• Be precise about what was and was not litigated. Do not tell anyone that "the PTAB invalidated the '140 patent" — it did the opposite. Conversely, do not concede that "claims 1–8 are valid"; the FWD holds only that Apple/Amazon failed to prove unpatentability on the grounds presented. Untested theories (eligibility, § 112, newly-found art, prior-use/on-sale art) remain open to a non-petitioner.
• Map estoppel before filing anything. Apple and Amazon are estopped as to grounds raised or reasonably raisable. If your client is in privity with either (supplier/ODM relationships, joint-defense agreements, funders with control) assume the estoppel attaches; if not, a fresh IPR on different primary art is the cleanest path, with the caveat that the Board has already shown it will credit Zentian's "reasonable expectation of success" and motivation-to-combine attacks on hardware-architecture combinations.
• Verify the expiry/term and maintenance-fee status on Patent Center, including any PTA and any terminal disclaimer in the continuation chain (the '140 issues from a 2019 continuation of a 2002-priority family). If the 2023-02-04 expiry is real, re-scope the case to past damages only.
• Trial-stage milestones: none are pending — both IPR2023-00037 and IPR2023-01197 are past FWD and in appeal, so there is no § 316(a)(11) one-year clock running and no new institution deadline to watch. The only live milestones are (a) the 2024-2208 decision and (b) whatever the stayed N.D. Cal. case (3:23-cv-02921-TLT) does if the stay lifts; a fact-discovery/claim-construction schedule will restart there on the district side.

I'll research the assignment chain and litigation context for US 10971140 before reconstructing the record.

Good litigation context gathered. Now let me pin down the assignment records, reel/frame data, and Zentian's corporate status.

US 10971140 — Ownership & Assignment Reconstruction

Analyst caveat up front: I could retrieve the existence and content of the recorded assignment (assignor, assignee, conveyance type, recording date) from Google Patents' legal-events mirror of USPTO assignment data, but I could not retrieve the reel/frame number or the correspondent of record for that entry from the USPTO Assignment Center within this session. I will not invent them. Every reel/frame slot below is marked NOT RETRIEVED rather than filled with a plausible-looking number.

---

Inventors

Pattern notes:

• Single-inventor patent. Unlike the sibling family (US 7,979,277 / US 10,839,789, which name Guy Larri, Mark Catchpole, Damian Kelly Harris-Dowsett, and Timothy Brian Reynolds), the '140 patent names Catchpole alone. Consistent with the specification, which is a single-technologist architectural disclosure.
• Inventor did NOT depart — no fire-sale precursor. Catchpole remained the named inventor across the entire Zentian portfolio (15 US patents per Patent Leaderboard, first granted 2009, most recent 2022), spanning the entire prosecution chain US 7,587,319 → 8,036,890 → 8,768,696 → 9,536,516 → 10,217,460 → 10,971,140. This is the inverse of the classic "all inventors gone within 12 months → portfolio dump" tell.
• Diversification after the speech-recognition work. Catchpole's later grants (US 11,457,868, US 11,406,012, US 11,272,580) are in conductive transfer / electrical-component transfer, co-invented with Paul Timothy Brook, Steven Paul Sutcliffe, Raymond Bungay and Melvyn Revitt. That is a different technical field entirely from Zentian's speech work — evidence the inventor moved on, not that the speech portfolio was abandoned.

---

Original assignee

Zentian Limited — a private company incorporated and registered in the United Kingdom, with its registered address given in its own pleadings as 27 Queens Close, St. Ives, Cambridgeshire, England, PE37 5QD (a residential address, which is the company's registered office, not a commercial facility). Cambridge, GB is the address on the face of the patent.

• Primary line of business: bespoke hardware for speech recognition — specifically a hardware accelerator for the acoustic stage (audio front end + distance calculation) intended to run alongside a software search stage.
• Did it ship a product embodying the claims? No evidence of any commercial product. Zentian's own complaints describe prototypes only ("the accelerator demonstrated a two-times speedup of the speech recognition software"), plus unsuccessful licensing/partnering conversations with Nuance, IBM, Hutchison 3G UK, Qualcomm and Apple — and it states plainly that "each speech recognition company expressed a desire to focus on software-only solutions." Apple "listened to Zentian's pitch and then ceased communication with the start-up."
• Current status: Operating as an IP-holding/assertion entity, not as a product company. Zentian is currently prosecuting and defending a coordinated multi-defendant campaign (see below). It is not in bankruptcy, not dissolved, and not acquired — no evidence of any of those.
• Data-integrity caveat (not an assignment issue): the '140 patent's citation lists on Google Patents are polluted with terahertz-imaging art (Terahertz Imaging of Breast Cancer; Tissue Classification Using Terahertz Pulsed Imaging; GB2347835, GB2380920, WO2000050859, US7174037) and even a laser-heterodyne interferometry patent (US5293213). None of these relate to speech recognition. Treat the citation list as unreliable; it does not affect the ownership chain.

---

Assignment timeline

Recorded assignments found for US 10,971,140 (Application 16/266,265): exactly one post-filing recordation, plus the original filing-stage assignment reflected in the patent front page.

• Executed date NOT RETRIEVED / recorded 2021-02-26 — Reel NOT RETRIEVED / Frame NOT RETRIEVED

  • Conveyance: Assignment (recorded on the USPTO register as "ASSIGNMENT OF ASSIGNORS INTEREST (SEE DOCUMENT FOR DETAILS)")
  • Assignor: CATCHPOLE, MARK (sole inventor)
  • Assignee: ZENTIAN LIMITED
  • Correspondent: NOT RETRIEVED from the Assignment Center. (I could not confirm the filing attorney/firm of record for this recording; I am therefore unable to run the repeat-correspondent test on the assignment chain. Do not read the litigation firms below as the assignment correspondent — they are a different role.)
  • Context: Internal chain-cleaning / inventory assignment — the sole inventor's rights formally confirmed into the original operating assignee roughly five weeks before grant of the '140 patent (2021-04-06) and roughly eleven months before the first assertion suits (2022-02-02). This is a title-hygiene recordation ahead of enforcement, not a change in beneficial ownership or a transfer to an asserter.
  • Recurrence flag: The same 2021-02-26 Catchpole→Zentian recordation pattern is the kind of entry that recurs across the other Zentian family members (the US 15/392,396 → 10,217,460, US 14/309,476 → 9,536,516, US 13/253,223 → 8,768,696, US 12/554,607 → 8,036,890 chain all name Zentian Limited as assignee on face). I could not verify correspondence across reels.

• Filing-stage assignment (on the face of the patent): Inventor Mark Catchpole → Zentian Limited, Cambridge (GB), as the pre-issuance assignment that supports the "(73) Assignee" field. The original 2003 PCT national-stage assignment for ancestor application US 10/503,463 (→ US 7,587,319) would sit under a separate reel/frame tied to that application, which is outside this patent's own assignment history.

No other recorded assignments. Specifically, no record of:

• any assignment out of Zentian Limited to any third party, LLC, or aggregator;
• any security agreement, merger, change of name, license, or release encumbering the patent;
• any security interest recorded against the patent.

If the Assignment Center holds additional entries I did not surface, that is a retrieval failure, not a finding of absence — verify directly at USPTO Assignment Center (search 10971140) or assignment.uspto.gov (search patent number 10971140).

---

Timeline diagram

timeline
    title Ownership and assertion of US 10971140
    2002 : GB priority application filed
         : Zentian Ltd named applicant
    2003 : PCT application filed
    2009 : Ancestor patent 7587319 granted
    2019 : Continuation 16266265 filed
    2021 : Assignment recorded Catchpole to Zentian Ltd
         : Patent 10971140 granted
    2022 : Suits filed vs Apple and vs Amazon
         : Apple IPRs filed against Zentian patents
    2023 : Amazon IPRs filed and joined
         : Second suit filed in Northern California
    2024 : PTAB final written decisions issued
         : Zentian appeals five decisions
    2026 : Federal Circuit rules on the appeals

---

NPE / troll-pattern signals

1. Shell-entity transfer — NOT PRESENT

There is no LLC in the chain. The patent has moved exactly once on the register, from the sole inventor to the original UK operating company (recorded 2021-02-26). No IP / Patents / Licensing / Holdings / Ventures suffix transferee exists. Caveat: Zentian's registered address (27 Queens Close, St. Ives — residential) and its status as a small private UK company would independently look shell-like, but the rubric requires the naming/address tell to be paired with a transfer to a licensing-only vehicle, and here there is no transfer at all. Not a finding.

2. Known asserter in the chain — NOT PRESENT on the enumerated lists / function-to-notify

Zentian Limited does not match any named entity on the supplied list (Acacia, Marathon, IV, IPNav, Wi-LAN, Mosaid/Conversant, Vringo, Pendrell, Innovatio, MPHJ, Lumen View, Round Rock, Document Generation Corp, Spangenberg). However, the assertion footprint is that of a serial commercial asserter, documented in the public record:

• Zentian Ltd v. Apple Inc., 6:22-cv-00122 (W.D. Tex., filed 2022-02-02)
• Zentian Ltd v. [Amazon.com, Inc.](/litigations/by-plaintiff/Amazon.com%2C%20Inc.), 6:22-cv-00123 (W.D. Tex., filed 2022-02-02)
• Zentian Ltd v. Apple Inc., 3:23-cv-02921 (N.D. Cal., filed 2023-06-14)
• Five patents asserted in one campaign: 7,587,319; 7,979,277; 10,062,377; 10,839,789; 10,971,140
• Twelve+ IPRs instituted against Zentian by Apple and Amazon (IPR2023-00033/00034/00035/00036/00037 and IPR2023-01192/01193/01194/01195/01197)

I did not retrieve an RPX or Unified Patents asserter-directory entry for Zentian, so I mark this unclear, not present. (Note: Google Patents' litigation panel labels the IPR2023-00037 data as sourced from "Unified Patents PTAB Data" — the petitioner of record is Apple Inc., not Unified Patents. Do not misread that label as Unified Patents asserting against Zentian.)

3. Repeat correspondent across the chain — UNCLEAR / NOT VERIFIABLE

The assignment-chain correspondent of record was not retrievable, so the recurrence test cannot be run on reels. Separately — and this is litigation counsel, not the recording correspondent — Zentian's appellate representation is Kayvan B. Noroozi (Noroozi PC, Los Angeles) argued at the Federal Circuit, with Jessica R. Bernhardt, Nevin M. Gewertz and Katherine E. Rhoades of Bartlit Beck LLP. Noroozi PC is a contingent-fee, patent-plaintiff-side boutique; Bartlit Beck's own attorney bio lists "Zentian, Ltd. v. Apple, Inc." and "v. Amazon.com, Inc. (W.D. Tex.)" as pending. Counsel choice is consistent with a monetization campaign rather than a product-defence posture, but I flag it as adjacent evidence only — it is not the assignment correspondent and I am not treating it as the recurrence signal.

4. Cascading transfers — NOT PRESENT

One link, one recording event. No chained LLCs, no cluster of assignees, no <24-month cascade, no shared correspondent address among transferees (there are no transferees beyond the original assignee).

5. Pre-litigation transfer — NOT PRESENT (timing is adjacent, not triggering)

The assignment was recorded 2021-02-26; the first suits were filed 2022-02-02 — a gap of roughly 11 months, outside the 6-month window. The recordation does fall in the narrow band before issuance (2021-04-06) and before enforcement, which is consistent with pre-enforcement title cleaning, but it is not a pre-litigation transfer and the assignee never changed. Not present on the literal test.

6. Bankruptcy fire-sale — NOT PRESENT

No Chapter 7/11, no administration, no insolvency proceeding surfaced for Zentian Limited. The patent was never sold out of an estate. Zentian still holds and is actively litigating it as of the 2024–2026 Federal Circuit appeals.

7. Privateering — NOT PRESENT

Privateering requires an operating company transferring to an NPE that asserts on its behalf. Here the asserter is the original operating entity — there is no upstream product company pulling strings, and no SEC-filing trail because Zentian is private and non-US.

8. Defensive aggregator — NOT PRESENT

The chain does not terminate at RPX, AST, LOT Network, Unified Patents, or OIN. Zentian remains the owner and is the plaintiff/appellant of record. The patent is live, not neutralized.

---

Verdict

NPE — moderate confidence

Justification. The assignment chain itself is an unusually clean one — a single recorded conveyance on 2021-02-26 from sole inventor Mark Catchpole to the original UK assignee Zentian Limited, with no shell LLC, no cascading transfers, no security interest, no bankruptcy, and no defensive-aggregator terminus. What drives the verdict is not the assignment structure but the absence of any product embodying the claims — Zentian's own complaints admit only prototypes and declined partnership pitches to Nuance, IBM, Hutchison 3G, Qualcomm and Apple — combined with a coordinated five-patent assertion campaign against Apple (6:22-cv-00122 W.D. Tex.; 3:23-cv-02921 N.D. Cal.) and Amazon (6:22-cv-00123 W.D. Tex.) filed in lockstep on 2022-02-02, staffed by contingent-fee plaintiff-side and litigation-boutique counsel. That is one strong signal (non-practice with serial assertion), which under the rubric yields moderate, not high, confidence — because every assignment-structure tell (shell transfer, cascading LLCs, pre-litigation transfer within 6 months, known asserter directory listing, privateering, fire-sale) is absent.

Verification links. Assignment records: USPTO Assignment Center (search 10971140) and assignment.uspto.gov (search patent number 10971140). Legal-events mirror used here: Google Patents US10971140B2 legal events. PTAB: IPR2023-00037 (Apple) and IPR2023-01197 (Amazon).

Two open items for a follow-up pass: (a) the reel/frame and correspondent of record for the 2021-02-26 entry — the single most diagnostic data point here, since a repeat-player recording attorney would upgrade signal 3 from unclear to present; and (b) confirmation of whether Zentian Limited remains in good standing at UK Companies House, which would rule bankruptcy/dissolution formally in or out rather than by absence of evidence.

I'll start by searching for the patent and its key prosecution/litigation context, then work through the cited references.

Let me verify the asserted IPR prior art and the appellate outcome, plus a few key cited references.

Prior‑Art Analysis — U.S. Patent No. 10,971,140 (Zentian Ltd.)

1. Verification of the patent (no substitute numbers)

Source: https://patents.google.com/patent/US10971140/en

Because the chain traces to a Feb. 4, 2002 priority date, pre‑AIA § 102/103 apply, and publication dates after 2002‑02‑04 matter only to the extent a reference has an earlier U.S. filing/priority date (§ 102(e) art) or is used for § 103.

---

2. Important caveat on "anticipation"

The task asks which claim(s) each reference "potentially anticipates under 35 U.S.C. § 102." I must be candid: pure § 102 anticipation requires a single reference disclosing every element of the claim as arranged. On the record I can review, no single reference cited on the face of the '140 patent discloses the complete combination of independent claim 1 — in particular (a) one or more clusters of processors each comprising a plurality of processors and an acoustic‑model memory, and (b) generating an "initial score" that is used to decide whether to continue processing to a final score by processing a larger amount of model data. That second limitation is the temporary‑lexical‑tree‑score / language‑model look‑ahead pruning concept from the specification (FIGS. 4 and 6), and the PTAB expressly found the claims not shown unpatentable over the art actually asserted against them (see § 5). So the cited references are best characterized as § 103 art, or as art bearing on individual dependent claims — not as clean § 102 anticipations of claim 1. I mark confidence accordingly below.

---

3. The citations appearing on the face of US 10,971,140

The reference list on the '140 patent ("Patent Citations (33)" / "Citations (32)") is reproduced below. Note an important data anomaly: the list is contaminated with references from an unrelated terahertz‑imaging family (Teraview Ltd./Teraprobe/Toshiba Res. Europe, and Klein's laser‑interferometry patent). These have no bearing on speech‑recognition claims and are flagged as non‑analogous. This appears to be an artifact of the Google Patents citation aggregation for this family; I would verify against the USPTO PatentCenter "References Cited" before relying on them.

3A. Speech/recognition‑relevant citations

(Item 27 intentionally omitted; see the non‑analogous list below — the face citations total 32/33 documents and I have not found a further speech reference beyond those listed.)

3B. Applicant's own earlier disclosure (not "prior art" in the § 102 sense)

• WO 2003/067572 A2 — Zentian Limited, "Speech recognition circuit using parallel processors," priority 2002‑02‑04, published 2003‑08‑14. This is the PCT parent of the '140 patent itself, not third‑party prior art. Also GB 2 384 901 B (Zentian, granted 2004‑04‑21) and GB 2 391 679 A (Zentian, 2004‑02‑11) are family members.

3C. Non‑analogous references (data contamination)

These are cited on the face of the family but belong to an unrelated terahertz‑imaging field; they would not be § 102 art for the speech‑recognition claims:

• US 5,293,213 A — Klein, "Utilization of a modulated laser beam in heterodyne interferometry" (1992‑08‑12 / 1994‑03‑08)
• WO 2000/050859 A1 — Teraprobe Ltd., "Method and apparatus for terahertz imaging" (1999‑02‑23 / 2000‑08‑31)
• GB 2 347 835 A — Toshiba Res. Europ., "Terahertz radiation imaging" (1999‑02‑23 / 2000‑09‑13)
• GB 2 380 920 A — Teraview Ltd., "Terahertz radiation imaging" (2001‑09‑12 / 2003‑04‑16)
• US 2003/0178584 A1 — Arnone, "Imaging apparatus and method" (2000‑02‑28 / 2003‑09‑25)
• US 7,174,037 B2 — Teraview Ltd., "Imaging apparatus and method" (2000‑02‑28 / 2007‑02‑06)

I flag these explicitly because a literal reading of the face citation list would otherwise suggest a prior‑art field that has nothing to do with the '140 claims. If you need this resolved, pull the "References Cited" box from USPTO PatentCenter for application 16/266,265.

---

4. Assessment: the most relevant prior art for the '140 patent

Ranked by relevance to the '140 claims:

1. US 6,526,380 B1 (Thelen, Philips) — most relevant architectural reference: parallel recognition engines, each with its own model (memory) and a controller distributing input and selecting results. Maps to the "clusters … each comprising a plurality of processors and an acoustic model memory" concept of claim 1, and to claims 7–8 (distinct memories / a model coupled to all processors of a cluster). Weakness: no "probability computed from acoustic model data in the shared cluster memory" and no "initial score gating further processing."
2. US 7,035,802 B1 (Matsushita) — most relevant algorithmic reference for claims 2–3: lexical trees, three‑state HMMs, state‑transition probability evaluation, Viterbi, and beam pruning.
3. US 5,983,180 A (Robinson, Softsound) and US 5,628,159 A (BBN) — tree‑structured finite‑state/HMM decoding, relevant to claims 1–3.
4. US 7,024,359 B2 (Qualcomm) — distributed recognition / feature‑vector handling, relevant to claims 1 and 5.
5. US 4,977,599 A (IBM) — HMM (incl. silence) modeling, relevant to claim 3.

Nothing in the face list directly addresses claim 4 (probability computed from a Gaussian mixture model and feature vectors) beyond general acoustic‑modeling context (e.g., Pioneer US 7,065,488), and nothing addresses the claim 1(e) "initial score → larger‑amount‑of‑model‑data" gating limitation.

---

5. The prior art that was actually asserted (and why it matters)

Separate from the face citations, the '140 patent was challenged in two IPRs, which is where the true § 102/§ 103 art sits:

• IPR2023‑00037 — Apple Inc. v. Zentian Ltd. (filed 2022‑12‑02; instituted 2023‑06‑12; Final Written Decision 2024‑06‑11). Challenged claims 1–8. Asserted art:
  • US 6,374,219 B1 ("Jiang") — "System for using silence in speech recognition," issued 2002‑04‑16 (continuation‑in‑part of Ser. No. 08/934,622). Discloses feature extraction → feature vectors/codewords, HMM phonetic models, prefix‑tree search engine, and beam pruning. Argued against claims 1(a)–(c), 2–4.
  • US 5,428,803 ("Chen") — parallel‑computing/shared‑memory reference relied on for the "clusters of processors, each cluster sharing a memory" architecture of claims 1(a)–(b), 7–8.
  • US 2001/0053974 A1 ("Lucke"), published 2001‑12‑20.
  • US 5,983,180 ("Robinson") — the Softsound reference also in the face list (claims 2–3).
  • US 5,036,539 ("Wrench"), issued 1991‑07‑30.
  • Outcome: The Board determined that Petitioner did not show any of claims 1–8 unpatentable ("Final Written Decision Determining No Challenged Claims Unpatentable"). This is strong, citable evidence that Jiang, Chen, Lucke, Robinson and Wrench — alone or in combination — do not anticipate the '140 claims.
• IPR2023‑01197 — Amazon Web Services, Inc. / [Amazon.com Services LLC](/litigations/by-plaintiff/Amazon.com%20Services%20LLC) v. Zentian Ltd. (2023): Ground 1 asserted claims 1–3, 5, 7–8 obvious over Jiang in view of Chen.

Appellate posture (flagging uncertainty): The '140 FWD was appealed to the Federal Circuit (docket 24‑2208). I have seen conflicting secondary reporting dated late September 2026 — one item describing an Apple win on the construction of "feature vector," another describing affirmance of an "unpatentable as obvious" holding. Those reports appear to relate to different Zentian patents in the family (the '277/'377/'789 group), and I cannot confirm with high confidence the disposition specific to the '140. The one fact I can ground is the PTAB FWD itself (no claims unpatentable), source: https://ptab.uspto.gov/ (IPR2023‑00037) and the FWD PDF.

---

6. Bottom line

• Most relevant prior art overall: US 6,526,380 (Thelen, parallel recognizers with per‑engine models) for the architecture, and US 7,035,802 / US 5,983,180 / US 5,628,159 for the HMM/lexical‑tree scoring of claims 2–3. In the litigation record, the art that actually mattered was Jiang (US 6,374,219) + Chen (US 5,428,803), with Lucke, Robinson and Wrench as secondary references.
• Anticipation (§ 102): I do not find a single face‑cited reference that discloses all elements of claim 1; the closest candidates (US 6,526,380; US 7,035,802) each miss at least the "initial score ↔ larger amount of model data" gating and/or the processor‑cluster‑with‑shared‑acoustic‑model‑memory combination. Treat these as § 103 references or as § 102 art against dependent claims 2–5.
• Caveats: (1) the face citation list is contaminated with unrelated terahertz references — verify against USPTO PatentCenter; (2) pre‑AIA § 102(e) governs several post‑2002 references (US 6,526,380, US 7,035,802, US 7,024,359, US 7,065,488, US 7,120,582); (3) the PTAB's 2024 FWD on the '140 is the strongest available evidence that the asserted art does not anticipate; (4) the Federal Circuit outcome on appeal (24‑2208) is not something I can state with high confidence.

If you want, I can next pull the exact "References Cited" list from USPTO PatentCenter for application 16/266,265 to eliminate the terahertz contamination and produce an element‑by‑element § 102/§ 103 claim chart for claims 1–8.

I'll ground the analysis in the references listed on the patent's Prior Art section. Let me verify the disclosures of the key references and confirm the IPR outcome.

Let me verify the specific IPR grounds (the references Apple actually asserted) and a couple of the listed prior-art references' disclosures.

Obviousness Analysis — U.S. Pat. No. 10,971,140 B2

Under 35 U.S.C. § 103(a) (pre‑AIA)

---

0. Scope, date, and data-quality flags (read first)

Three things must be stated before the analysis, because they materially affect reliability.

1. Statutory framework. The application 16/266,265 was filed 2019‑02‑04, i.e. after the AIA's 2013 date. However, it is a straight continuation in an unbroken chain running to PCT/GB2003/000459 and GB 0202546.8 (2002‑02‑04), and nothing in the granted claims introduces new matter. Every claim therefore has an effective filing date of 2002‑02‑04, so pre‑AIA § 102/103 governs. The critical date for § 102(b) art is on or about 2002‑02‑04 (one year before the 2003‑02‑04 U.S./PCT filing).

2. Date discrepancy (flagged, per instructions). My task header states the current date is April 26, 2026, while the environment date is 2026‑09‑30, and the previously generated sections (and the live results I retrieved) reference events after 2026‑04‑26 — e.g. the Fed. Cir. oral argument in Apple v. Zentian, No. 2024‑2208, on 2026‑06‑04, and the Sept. 25, 2026 nonprecedential decisions in the sibling appeals. I cannot reconcile these; I treat the 2026‑09‑30 environment date and the retrieved sources as ground truth, and I do not rely on the April 26 date.

3. What "the Prior Art section of this page" actually contains, and its limits.

   • The Google Patents page (https://patents.google.com/patent/[US10971140](/patent/US10971140)/en) lists ~33 patent citations (examiner + third‑party), 11 non‑patent citations, "Families Citing this Family" (29), and "Cited By" (12/6).
   • Several list entries are not prior art to a 2002‑02‑04 priority date: WO2003067572A2 (this patent's own PCT publication), US20080255839A1 (Zentian, 2004), and US7899669B2 (2005). These are the applicant's own later filings / family members, not § 102 art.
   • Several are citation-record artifacts — the terahertz‑imaging papers and patents (Woodward, Wallace, Cole, Fitzgerald; WO2000050859A1; GB2347835A; GB2380920A; US7174037B2; US20030178584A1) and US5293213A (laser interferometry) — unrelated to speech recognition. (The prior section already flagged this; I confirm it and exclude them.)
   • Critically: the art the Board actually adjudicated for this patent — Jiang, Chen, Lucke, Wrench, and Robinson — is largely not on this page's prior‑art list. (Robinson = US 5,983,180, which is listed, as cited by the examiner.) Any § 103 theory built only from the page's list is therefore my own construction, not the Board's, and must be weighed against the actual outcome. I flag this as the single most important limitation of this analysis.

---

1. The claim to be tested

Claim 1 (sole independent claim) — decomposed into limitations:

Claim 1 is drafted functionally and broadly. It does not recite lexical trees, a results memory, a controller, or a language model — all of which are central to the 2002 specification and to the abstract. It recites, in substance, a clustered, shared‑memory, parallel hardware platform (1a–b) that performs a coarse‑first/gated computation (1c/1e). The claim‑construction fight in the family has been over "feature vector" (see § 6), and the Board itself noted the parties pressed "implicit claim constructions."

---

2. Mapping the listed prior art to claim 1

2.1 Limitation 1(a) — clusters of processors + acoustic model memory

• US 6,526,380 B1 (Thelen et al., Koninklijke Philips; filed 1999‑03‑26; § 102(e) date qualifies). Discloses a plurality of N large‑vocabulary speech recognizers, each associated with a respective, different recognition model targeted to a part of a huge vocabulary, plus "a controller operative to direct the input pattern to a plurality of the speech recognizers and to select a recognized word sequence." It expressly states recognition may be done "using separate recognition engines each having own processing resources," and that the models are hierarchically arranged. URL: https://patents.google.com/patent/[US6526380B1](/patent/US6526380B1)/en. This reads directly on "clusters of processors" + "acoustic model memory" (each recognizer's own model store) + a distributing controller.
• US 7,034,802 B1 (Matsushita; filed 2000‑07‑31; § 102(e) qualifies). A lexical‑tree speech recognizer using HMMs; relevant background/architecture. URL: https://patents.google.com/patent/[US7035802B1](/patent/US7035802B1)/en.
• NPL — Glinski et al., "Spoken Language Recognition on a DSP Array Processor," IEEE Trans. Parallel & Distributed Systems (1994), 697–703 and Chatterjee et al., "Connected Speech Recognition on a Multiple Processor Pipeline," ICASSP 89, 774–777 and Chung et al., "A Parallel Phoneme Recognition Algorithm Based on Continuous Hidden Markov Model," IPPS/SPDP 1999, 453–457. These are the page's NPL entries that establish parallel/DSP‑array/pipelined hardware for HMM speech recognition — i.e. that placing a plurality of HMM processors behind shared model storage was a known architecture by the priority date.

Assessment: 1(a) is met by any of US 6,526,380 (alone or in view of the array/pipeline NPL), with US 7,035,802 supplying the HMM lexical‑tree memory structure.

2.2 Limitation 1(b) — each processor computes a probability from the acoustic model data

• Thelen (US 6,526,380) — each recognizer "evaluat[es]" input against its recognition model (acoustic model + vocabulary); the recognizer of its Fig. 1 shows the classic spectral‑analysis → unit‑matching → decision pipeline, i.e. probability computation against stored acoustic models.
• US 4,977,599 A (IBM; 1985‑05‑29) — speech recognition using a set of Markov models including silence models; teaches HMM probability/state computation.
• US 4,980,918 A (IBM; 1985‑05‑09) — efficient storage and assembly of phonological graphs.
• US 7,035,802 (Matsushita) — HMM state/transition probability computation over Gaussian/probability densities, with feature vectors (its Fig. 1 discussion of the "collection of probabilities … associated with the states … and … transition").

Assessment: 1(b) is squarely disclosed. Computing an acoustic probability against a stored model is the definition of HMM‑based ASR.

2.3 Limitations 1(c)/1(e) — the crux: "initial score" used to gate a larger "final score" computation

This is where the claim lives or dies, and the Board said exactly that (the challenged limitations were styled 1(d)/1(e), "generate an initial score" / "used to determine whether to continue … via processing a larger amount of model data").

The listed prior art teaches the concept from three independent directions:

• NPL — Deshmukh, Ganapathiraju & Picone, "Hierarchical Search for Large‑Vocabulary Conversational Speech Recognition," IEEE Signal Processing Magazine 16(5), 84–107 (Sept. 1999) — the very paper cited in the '140 patent's own Background. DOI 10.1109/79.790985. Its abstract states, verbatim: "multi‑pass systems that perform a quick search using a simple system, and then rescore only the N‑best resulting hypotheses using better models are very popular for real‑time implementation" and "fast‑matching strategies within the acoustic model are important … vector‑quantization‑like approaches that enable one to compute only a small number of Gaussians per frame." It also describes look‑ahead and beam pruning: "heavier pruning … Look‑ahead and N‑best strategies at all levels of the system." This is precisely "generate an initial score … used to determine whether to continue … via processing a larger amount of model data." URL (abstract mirror): https://www.mendeley.com/catalogue/badd62e3-6bf9-3b27-a071-8a88e3a111d0/
• US 7,035,802 B1 (Matsushita) — states the disclosed lexical‑tree algorithm "can be used for isolated word recognition, or as a first pass fast match for continuous speech recognition." A first‑pass fast match is, by definition, a cheap initial score used to gate a fuller, more expensive computation.
• US 5,710,866 A (Microsoft; 1995‑05‑26) — "speech recognition using dynamically adjusted confidence measure": computing a confidence/metric and using it to govern subsequent processing.
• US 4,977,599 / 5,988,180 (Robinson) / US 5,994,215‑type tree‑pruning art on the list (US 5,983,180 = Softsound "finite state sequence models organized in a tree structure"; US 5,995,930 = Philips vocabulary tree) supply pruning of a search to decide what to keep processing — the functional equivalent of "whether to continue."

Assessment: 1(c)/1(e) — the arguably novel twist — is disclosed or at least rendered obvious by the prior art that the applicant itself cited as background (Deshmukh), combined with the "first‑pass fast match" teaching of US 7,035,802.

---

3. The combinations, with explicit motivation

Combination A (primary): Thelen (US 6,526,380) + Deshmukh (NPL) + Matsushita (US 7,035,802)

• Thelen supplies 1(a) and 1(b): plural recognizers = plural processors, each with its own recognition model = acoustic model memory; controller distributes input; engines may be separate hardware.
• Matsushita supplies the HMM/lexical‑tree probability machinery (1b) and the express "first pass fast match" teaching (1c/1e).
• Deshmukh supplies the hierarchical/coarse‑to‑fine, multi‑pass, fast‑match, look‑ahead gating (1c/1e) that the '140 patent's own Background adopts.

Motivation (KSR‑sufficient):

1. Same field, same problem, same art the applicant acknowledged. Deshmukh is cited in the '140 specification itself as describing "the general concepts of large vocabulary speech recognition." A reference that the inventor's own background section treats as describing the state of the art is a fortiori combinable.
2. Articulated, known engineering goal: real‑time LVCSR. Deshmukh states the whole point of multi‑pass/fast‑match is to reach "near‑real‑time speed without a great sacrifice in word error rate." Thelen states parallel recognizers are used so "the user does not experience a significant delay."
3. Predictable result / finite set of known techniques: combining (i) parallel recognition engines and (ii) two‑stage coarse‑to‑fine gating is a combination of familiar elements performing their known functions with no change in their respective principles of operation → KSR, 550 U.S. at 417–421.
4. Design‑choice + "obvious to try": Where the prior art teaches a range of pruning/beam‑width strategies (Deshmukh: "the optimal values of the various beam widths … need to be derived … empirically"), selecting an initial‑score gate is a predictable configuration, not an inventive leap.

Combination B (secondary): Parallel‑hardware NPL (Glinski / Chatterjee / Chung) + Thelen + Deshmukh

• Glinski's DSP array processor and Chatterjee's multiple‑processor pipeline establish that distributing HMM ASR across an array/pipeline of processors was known. Chung's "Parallel Phoneme Recognition Algorithm Based on Continuous Hidden Markov Model" explicitly ties parallel processing to continuous‑HMM probability computation.
• Motivation: speed‑up of the computationally intensive Viterbi search, the express object of the '140 patent ("It is an object of the present invention to provide an improved circuit which can perform parallel processing of speech parameters"). Achieving a known goal by a known hardware arrangement = obvious.

Combination C (for 1(a)'s "clusters" specifically): Thelen's hierarchy + Glinski/Chung

Thelen's "recognition models … hierarchically arranged" and "separate recognition engines each having own processing resources" map onto "clusters." Motivation: Thelen itself gives it — start with generic models, escalate to more specific ones ("If a certain generic model proves to provide good recognition result, more specific models can be tested"), which is itself a gating/continuation decision from an initial to a final result.

---

4. Dependent claims 2–8

Motivation for the dependent claims: each adds a conventional, well‑understood implementation detail (HMM topology; GMM scoring; buffering; a master/controller processor; partitioned model memory) whose use with the base combination is a predictable design choice with no new principle of operation (KSR). For claim 6 specifically, Thelen's own controller provides the motivation ("direct the input pattern to a plurality of the speech recognizers").

---

5. A POSA profile and the level of ordinary skill

The Board and the parties converged on a POSA with an advanced degree (or equivalent experience) in CS/EE and ≥2–3 years in speech recognition/signal processing. Such a person, by 2002, would know: HMM/Viterbi decoding, beam pruning, fast‑match/look‑ahead, Gaussian scoring, and the use of DSP arrays and parallel recognizers. That profile makes each above mapping within ordinary skill, which is the § 103 touchstone.

---

6. The counter‑case — and why it dominates this analysis

I must state plainly: the art I can build from the page's "Prior Art" list is not the art the Board actually tested, and the Board rejected that art.

• PTAB, IPR2023‑00037, Final Written Decision, 2024‑06‑11 (Paper 34): "Determining No Challenged Claims Unpatentable" as to claims 1–8. FWD PDF: https://www.docketalarm.com/cases/PTAB/IPR2023-00037/Apple_Inc._v._Zentian_Limited/docs/06-11-2024-Board/Final_Written_Decision__original-34-Judgment_Final_Written_Decision_.pdf
• The asserted grounds were: Jiang + Chen (claims 1–3, 5, 7, 8); Jiang + Chen + Lucke; Jiang + Chen + Robinson and + Lucke + Robinson (claim 4); Jiang + Chen + Wrench (claim 6). Institution decision: https://www.docketalarm.com/cases/PTAB/IPR2023-00037/Apple_Inc._v._Zentian_Limited/docs/06-12-2023-Board/Institution_Decision__Grant-10-Granting_Institution_of_Inter_Partes_Review_35_USC_%C2%A7314.pdf
• Why it failed (per PO's demonstratives and the oral‑argument transcript): the fight was over limitations 1(d)/1(e) — exactly my 1(c)/1(e). Patent Owner argued (i) no reasonable expectation of success in substituting Chen's clustered parallel hardware into Jiang's PC‑based recognizer, and (ii) no motivation to combine. PO's demonstratives: https://www.docketalarm.com/cases/PTAB/IPR2023-00037/Apple_Inc._v._Zentian_Limited/docs/03-04-2024-Patent_Owner/Exhibit-2022-Patent_Owners_Demonstratives.pdf. At the Fed. Cir., Zentian framed the gap as: Jiang lacks clusters; the pruning algorithm therefore "must be combined into the clustered processor architecture of Chen," and the petition "was entirely silent" on how to do so. (Oral argument audio/transcript, Apple v. Zentian, No. 2024‑2208: https://www.courtlistener.com/audio/[105305](/patent/105305)/apple-inc-v-zentian-ltd/)
• IPR2023‑01197 (Amazon) was instituted and joined to the ‑00037 proceeding on 2024‑01‑10.

Implications for any § 103 theory:

1. The "circuit"/hardware‑combination gap is real. The failed petition foundered precisely on how to place a known algorithm into a clustered parallel hardware architecture. My Combinations A–C carry the same structural weakness: US 6,526,380 and the NPL array references are systems/methods, and bridging "software recognizer" → "cluster of hardware processors sharing an acoustic model memory" requires a concrete, non‑hindsight rationale — not merely "combine for speed."
2. The Federal Circuit appeal (No. 2024‑2208) is unresolved in my sources. I located the 2026‑06‑04 oral argument but no disposition for 2024‑2208. Do not treat the appeal as decided. The Sept. 25, 2026 nonprecedential decisions (Fed. Cir. Blog: https://fedcircuitblog.com/2026/09/25/opinions-orders-september-25-2026/) and the "feature vector includes codewords" holding concern U.S. 10,062,377 / 10,839,789 (appeals 24‑2204/2205/2207), not the '140 patent — consistent with, and reaffirming, the prior section's disambiguation.
3. Secondary considerations / nexus: Patent Owner's success (and the 100% survival of claims 1–8 against the asserted art) is consistent with — though not proof of — non‑obviousness. A rigorous petitioner would also have to address commercial success/praise and any unexpected results from the clustered sharing of model memory (the spec argues it avoids "bandwidth limited by the interface to the memory," and "avoids any bottleneck").

---

7. Bottom line

• On the face of the page's listed prior art and the patent's own Background (Deshmukh), claim 1 is, in my view, prima facie obvious: US 6,526,380 (Thelen) discloses parallel recognizers each with its own model memory plus a distributing controller; US 7,035,802 (Matsushita) supplies HMM probability computation and an express "first‑pass fast match"; and Deshmukh — cited by the applicant itself — supplies the initial‑score/coarse‑to‑fine gating of limitations 1(c)/1(e). Claims 2–8 are conventional implementation details with motivating teachings in the same references.
• But the defensible conclusion is that any such theory is fragile. The one tribunal that tested this exact claim (PTAB, IPR2023‑00037) held all of claims 1–8 not unpatentable, and the failure was localized to precisely the two limitations I flagged as the crux (1(d)/1(e)) — with the Board and Patent Owner faulting the petitioner for an incomplete hardware‑combination showing. A § 103 attack built from the page's list would have to overcome the same "how do you actually build it into a clustered circuit" problem, and would face the unverified (as to me) but potentially dispositive Fed. Cir. 2024‑2208 outcome.

Confidence: high on the references' bibliographic data and on the IPR/§ 103 posture; medium/low on any conclusion of actual invalidity, because (a) I could not retrieve the full FWD claim‑by‑claim reasoning, (b) the art actually adjudicated (Jiang, Chen, Lucke, Wrench) is not on this page, and (c) the appeal outcome is unconfirmed.

Unverified items (explicit):

1. Disposition of Fed. Cir. No. 2024‑2208 (appeal of the '140 IPR) — none found.
2. Full text/reasoning of the 2024‑06‑11 FWD beyond the judgment line and party summaries.
3. Whether the asserted Jiang/Chen/Lucke/Wrench references independently read on claim 1 (I did not locate their full disclosures).
4. Current USPTO fee/status for 16/266,265 beyond Google Patents' "Expired – Fee Related."