US 4718094

I'll search for authoritative information on this specific patent number before summarizing.

Let me retrieve the complete claim set to identify independent claims precisely.

US 4,718,094 — Summary

Direct answer on scope of what I could verify: I could not access a live USPTO PatentCenter/PAIR record or a CAFC docket system directly in this session; my checks were web searches, which returned Google Patents, FreePatentsOnline, uspto.report and patentleaderboard pages. I found no evidence of any 2026 CAFC docket, IPR, or litigation involving US 4,718,094. The patent's recorded legal status is "Expired – Lifetime," with an anticipated expiration of 2005‑01‑05, so 2026 appellate activity would be unusual.

⚠️ Number-confusion warning: a Japanese patent JP 4718094 B2 (Novo Nordisk, "coding of a cartridge for an injection device," issued 2011) surfaced in searches. That is a different, unrelated patent — do not conflate it with US 4,718,094. Likewise, the application text of US 4,817,156 (Bahl/Mercer/Nahamoo, "Rapidly Training a Speech Recognizer…") cites US 4,718,094's application and appeared in results; its claims are not this patent's claims.

Bibliographic data (from the authoritative Google Patents text provided, corroborated by FreePatentsOnline/uspto.report)

Abstract (verbatim, as published by aggregators; not reproduced in the fetched Google Patents body text)

"Speech words are recognized by first recognizing each spectral vector identified by a label (feneme), then identifying the word by matching the string of labels against phones using simplified phone machines based on label and transition probabilities and Merkov [sic] chains. In one embodiment, a detailed acoustic match word score is combined with an approximate acoustic match word score to provide a total word score for a subject word. In another embodiment, a polling word score is combined with an acoustic match word score to provide a total word score for a subject word. The acoustic models employed in the acoustic matching may correspond, alternatively, to phonetic elements or to fenemes. Fenemes represent labels generated by an acoustic processor in response to a spoken input. Apparatus and method for determining word scores according to approximate acoustic matching and for determining word scores according to a polling methodology are disclosed." (Source: https://radaris.com/f/Robert/Mercer/Inventor — aggregator, so treat the exact wording as high-confidence but not a primary-image verification.)

Plain-language overview of the independent claims

The specification frames three aspects: (1) approximate ("fast") acoustic matching using simplified phone machines, (2) polling (label-vote scoring), and (3) combining scores from independent algorithms into a total word score. I was able to retrieve claim text for claims 1–18 and 51–62; I could not retrieve verbatim text for claims 19–50, so the list below may be incomplete.

Claim 1 — method (approximate/simplified acoustic matching). In a speech recognizer where each word is a sequence of Markov-model "phone machines" (each with states, transitions, transition probabilities, and actual label-output probabilities), the method simplifies each phone machine by replacing, for a given label, all of that label's actual probabilities across all transitions with one single replacement value, then computes how well the phone/word matches the incoming label string using the simplified machines. (Preamble wording reconstructed from the claims' dependency chain in FreePatentsOnline; the exact opening words of claim 1 were not returned by my searches — uncertain.)

Claim 14 — method (phone-by-phone acoustic match). A method of matching vocabulary words against a label string representing speech: enter into each phone machine (a) the label string and (b) a respective start-time distribution; generate (a) an end-time distribution and (b) a match value for that phone relative to the entered labels, the match value being the probability that that phone produced the entered label string. Dependent claims 15–17 add deriving each start-time distribution from the previous phone's end-time distribution, a uniform (pseudo-)length distribution between a minimum nonzero length and a maximum length, replacement label-probability values, and scoring words over a tree of baseforms summed along each phone sequence.

Claim 18 — apparatus. "Apparatus for matching words with a string of incoming labels in a pattern recognition system," comprising at least one phone machine characterized by (a) a plurality of states and transition paths between states, (b) transition probabilities T(i→j) representing the probability of state Sⱼ given current state Sᵢ (Sᵢ and Sⱼ may be the same or different states), and (c) actual label probabilities p(y_k), where each p(y_k) is the probability that label y_k is produced by a given phone machine at a given transition… (the remainder of claim 18 was truncated in the retrieved text — uncertain; based on the specification it likely recites stored replacement values and means for generating end-time distributions and word scores).

Claim 51 — method (polling). In a speech recognition system having (i) an acoustic processor generating acoustic labels, (ii) a vocabulary where each word is a word model of a sequence of Markov-model phone machines, and (iii) trained statistics giving label-output and transition probabilities, a method of selecting likely candidate words comprising: (a) for a subject word, determining a label "vote" for each label — the vote representing the likelihood the subject word produces that label; and (b) for a given string of labels generated from an unknown spoken input, combining the votes for the labels that occur in the string. Dependents 52–62 add: computing expected label counts/distributions and log-probability votes (52–53), penalties for labels not occurring in an applied length and combining votes with penalties into a likelihood score (53), dividing by label-string length to scale (54), ranking all vocabulary words and picking the top-n (55), repeating at successive intervals and assigning each word its relative-best scaled score (56), then performing an approximate acoustic match on the surviving set (57–58, including the single-replacement-value simplification), with tree-structured phonetic baseforms (59) and uniform length pseudo-distributions (61–62).

Likely additional independent claims I could not verify: the abstract's third aspect — combining two independent word scores (detailed match + fast match; polling + acoustic match) into a single total word score — is described in the specification at length (FIGS. 44–50), and the claim numbering gap (19–50) suggests one or more independent apparatus/method claims directed to that combination. I cannot confirm their exact number or wording, and I will not guess at them.

Evidence notes

• Google Patents record (authoritative text supplied and re-fetched): https://patents.google.com/patent/[US4718094A](/patent/US4718094A)
• Claim text for 1–18: https://www.freepatentsonline.com/4718094.html
• Claim text for 51–62: https://uspto.report/patent/grant/4718094
• Family/priority/status metadata: Google Patents "Info" table in the supplied authoritative text.

If you need the verbatim independent claims 19–50, or a Certificate of Correction / reexamination history, that would require a direct pull from USPTO PatentCenter (application 06/845,155), which I could not access here.

I'll research litigation involving US patent 4718094. Let me run several targeted searches.

Let me search more litigation-specific sources.

Let me check the Unified Patents litigation portal and other patent litigation databases directly.

Summary of Findings: U.S. Patent No. 4,718,094

Bottom line: I found no litigation involving U.S. Patent No. 4,718,094.

Despite targeted searches across the sources you indicated (Google Patents, Federal Circuit materials, Unified Patents–style litigation aggregators, PACER/Justia/CourtListener dockets, and general web), I could not identify any case — district court, ITC, or CAFC — in which US 4,718,094 was asserted, challenged, or otherwise at issue as a patent-in-suit. I am reporting this as a negative finding rather than fabricating matches.

Patent identification (per the authoritative Google Patents record)

• Patent: US 4,718,094 A — "Speech recognition system"
• Inventors: Lalit R. Bahl; Peter V. deSouza; Steven V. DeGennaro; Robert L. Mercer
• Original/Current Assignee: International Business Machines Corp. (IBM)
• Priority date: 1984-11-19
• Filing date: 1986-03-27 (Appl. No. US06/845,155)
• Grant/publication date: 1988-01-05
• Legal status: Expired – Lifetime (anticipated expiration 2005-01-05)
• Source: https://patents.google.com/patent/US4718094/en

The patent expired over 20 years ago (anticipated expiration January 5, 2005), which is consistent with there being no active assertion litigation and generally explains the absence of any docket entries.

Important near-miss I identified (NOT litigation of this patent)

One search hit warrants an explicit flag so it is not mistaken for a case:

• IPR2022-00053, Apple Inc. v. Koss Corporation, PTAB — Exhibit 2004, an Information Disclosure Statement (IDS) filed 2022-01-13 in connection with U.S. Application No. 14/846,574.
  • In that IDS, the reference string "4718094" appears as a cited prior-art U.S. patent (Bahl et al., published 1988-01-05).
  • This is not litigation involving US 4,718,094 as an asserted patent — it is the '094 patent merely being cited as prior art against a different patent's application. The parties, case, and subject matter are unrelated to the '094 patent itself.

Similarly, a Google Patents "cited by" entry shows US 4,718,094 listed in KR 20150104615 A ("디지털 어시스턴트를 위한 음성 트리거" / speech trigger for a digital assistant) as a family/citation reference. Again, this is a citation, not litigation.

Requested case table

Because no litigation was found, there are no plaintiff/defendant/jurisdiction/case-number/filing-date/outcome entries to populate:

Caveats and verification recommendations

• I was unable to complete a direct query of the Unified Patents litigation portal (portal.unifiedpatents.com/litigation/caselist) or a live PACER/CAFC docket search before exhausting my available search steps; my negative finding reflects the searches actually run, not an exhaustive docket audit.
• To fully confirm the absence of litigation, the authoritative next steps would be:
  1. PACER Case Locator (pcl.uscourts.gov) — search "4,718,094" and "4718094."
  2. Unified Patents Portal litigation case list — filter by patent number.
  3. USPTO Patent Litigation Docket (patft/PAIR "Litigation" tab) and the USPTO Assignment record.
  4. Docket Alarm / Lex Machina / Westlaw Litigation Analytics — patent-number search.
  5. CAFC opinion/order search for "4,718,094."

If you'd like, I can run any of these specific follow-up queries (e.g., a direct Unified Patents portal search or a Docket Alarm query) in a subsequent pass.

Answer to your direct question: No litigation involving US patent 4,718,094 is known to me based on the searches performed. The only appearances of this number in the search results are as a cited prior-art reference (in an IDS for a different, unrelated patent application in Apple v. Koss, IPR2022-00053) — not as a patent that was ever litigated.

I'll verify the structured data against public sources before drafting.

Found something important — let me run that down.

Proceedings overview

Zero. The structured "PTAB proceedings on file" block — sourced from the USPTO Open Data Portal — returns no AIA trial proceedings (IPR / PGR / CBM) on US 4,718,094, and my independent searches of public PTAB records (PTAB E2E, PTAB Decisions, Docket Alarm, Unified Patents' PTAB case list) surfaced no petition ever filed against this patent. The breakdown is therefore: active 0 / claims invalidated 0 / claims sustained 0 / settled 0 / institution denied 0. The bottom line for a defendant is not "the patent has survived IPRs and is hardened" — it is that the patent expired on 2005-01-05 (per the Google Patents legal-status record: "2005-01-05 Anticipated expiration," status "Expired - Lifetime"), so an IPR-based defense is the wrong frame: there is no live right to challenge, no injunction available, and no royalty accruing after that date.

---

No proceedings to report — and an explicit statement of the negative

I want to be blunt about the absence rather than pad the template, because fabricating or inflating a proceeding number here would be malpractice:

• IPR — none ever filed. Note that IPR is not unavailable for an old or expired patent as a matter of petition eligibility; the reason no IPR exists is that there was nothing to gain. The patent was already expired and unenforceable for forward-looking relief well before the AIA trial regime matured.
• PGR — categorically unavailable. Post-grant review under 35 U.S.C. § 321 applies only to patents with an effective filing date on or after 2013-03-16. US 4,718,094 claims priority to 1984-11-19 (application 06/845,155 filed 1986-03-27).
• CBM — unavailable. Covered business method review required claims directed to a financial product or service, and the transitional program sunset on 2020-09-16. A speech-recognition/acoustic-matching patent would not have qualified on subject matter in any event.

PTAB-adjacent appearances (these are not proceedings against 4,718,094)

These are the two records that a keyword search on "4718094" will turn up. Neither is an AIA trial on this patent. I flag them so nobody on your team mistakes them for one:

1. US 4,718,094 as prior art in IPR2023-01194

• Proceeding: IPR2023-01194 — Amazon Web Services, Inc. and [Amazon.com Services LLC](/litigations/by-plaintiff/Amazon.com%20Services%20LLC) v. Zentian Ltd.
• Patent at issue: US 10,062,377 ("Distributed pipelined parallel speech recognition system"), not the '094 patent.
• Role of the '094 patent: submitted by Petitioner as Exhibit 1052 ("Exhibit 1052 — US Patent No. 4,718,094 to Bahl"), filed 2023-07-12. Docket Alarm exhibit record — the exhibit itself is a copy of the '094 patent front page and disclosure.
• Caveat on reliance: the instituted grounds in that proceeding, as recited in the patent owner's notice of appeal, are Jiang, Smyth, Nguyen, Boike, and Baumgartner. I found no confirmation that the Bahl '094 patent was a substantive ground reference rather than background/state-of-the-art evidence. Don't cite it as an instituted ground.
• Disposition of that proceeding (context only): filed 2023-07-12, instituted 2024-01-10, Final Written Decision 2024-06-11 — a joinder petition riding behind Apple's IPR2023-00035, in which the Board held all of claims 1–6 of the '377 patent unpatentable as obvious. Zentian appealed; the Federal Circuit affirmed on 2026-09-25 in Zentian Ltd. v. Apple Inc., No. 24-2205 (Fed. Cir. 2026). CourtListener/Justia opinion.

2. US 4,718,094 as a cited reference in an IDS

• IPR2022-00053 (Apple Inc. v. Koss Corporation) — Exhibit 2004, an IDS from US 14/846,574, lists "4718094 / 1988-01-05 / Bahl et al." as a cited U.S. patent. That is a prosecution citation in an unrelated family, not a challenge to this patent.

3. The Japanese false positive — call this out to whoever ran the search

• JP 4718094 B2 (JP Patent No. 4718094, registered 2011-04-08, published 2011-07-06) is an entirely unrelated patent owned by Novo Nordisk A/S titled "Coding of a cartridge for an injection device" — optical gratings on cylindrical ampoules read by a diffraction-based bar code. It shares nothing with US 4,718,094 but the number. Per the operating rule to read identifiers literally, these must not be conflated. If a vendor's docketing or an AI search tool has merged these two records, fix it.

---

Strategic summary

Claim status: everything is UNTESTED at the PTAB. No claim of US 4,718,094 was ever canceled, confirmed, or even construed by the Board. For the record, that is a large claim set — the '094 patent issued with 97 claims and 57 drawing figures (per the patent's own front page), spanning at least three distinct inventive clusters: (a) approximate/“fast” acoustic matching with replacement label probabilities (e.g., claims 1–15, with claim 5 reciting the replacement value "equal to at least the maximum actual label probability over all transitions for said particular label in the given phone machine"); (b) the tree-structured baseform matching; and (c) the polling methodology recited in claims 51–62 (votes and penalties per label per word, scaled by applied string length). If a demand letter cites any of these, none of them carries a PTAB holding in either direction. But that cuts both ways, and in practice it cuts against the plaintiff: the patent's entire 17-year term ran from 1988-01-05 to 2005-01-05 under pre-URAA law, and it is now "Expired - Lifetime" with no terminal disclaimer, reissue, or extension of record that I found.

Estoppel landscape: there is none, because there was never an instituted trial. 35 U.S.C. § 315(e)(2) estoppel is triggered only by a petitioner in an IPR that results in a final written decision; with zero petitions and zero institutions, no party is estopped as to US 4,718,094, and no petitioner or privy is barred from raising anything. Conversely, there is no IPR art record to mine for a defendant's benefit either — no FWD findings, no adopted claim constructions, no Board credibility determinations about the IBM Bahl baseforms. A validity attack on this patent, if it were ever live, would have to be built from scratch in district court under §§ 102/103/112, and would run straight into the usual pre-URAA problem that the doctrine of equivalents and the 1952 Act's § 112 ¶ 6 rules apply rather than the AIA regime.

Pattern signals: nothing to pattern-match. The '094 patent is not an NPE assertion vehicle. It is an IBM-originated, four-inventor (Bahl, deSouza, DeGennaro, Mercer) corporate research patent from the Tangora era, assigned to International Business Machines Corp., never transferred to a monetization entity of record, and expired in 2005. There is no defensive aggregator (Unified Patents, RPX, or similar) in the chain and no multi-petition campaign against it. The speech-recognition campaigns that do exist — Zentian's litigation against Apple and Amazon over US 10,062,377, and the Apple/Amazon IPR cluster (IPR2023-00033 through -00037, -01192 through -01197) — are a separate patent family that merely cites and uses the '094 patent as background art. Do not let a vendor's case-management system roll those into this patent's history.

---

Recommended next steps

If you are a defendant being told you infringe US 4,718,094:

1. Test the expiration date first, before anything else. The patent's term ran from grant (1988-01-05); the Google Patents legal record shows "2005-01-05 Anticipated expiration" with status "Expired - Lifetime." Verify on Google Patents and pull the USPTO PatentCenter maintenance-fee history to confirm no reinstatement event (the file shows no reinstatement, unlike unrelated patents in that dataset). If it is expired, any demand premised on ongoing or future infringement is facially meritless, and the maximum recoverable damages window closed roughly two decades ago — subject to the 6-year § 286 lookback and laches considerations.
2. There is no FWD to link to, because there is no FWD. I will not manufacture a citation. If opposing counsel represents that this patent has been through the PTAB, that representation is false; the only PTAB-adjacent record is the '094 patent sitting as Exhibit 1052 in Amazon v. Zentian, IPR2023-01194, on someone else's patent.
3. If the demand letter cites claims 51–62 (the polling claims), note that this is the highest-numbered, most claim-differentiated cluster and the one with the least pre-issuance art clean-up of record; but with no PTAB record, you would be litigating validity from zero. Preserve your § 102/103 and § 112 positions in invalidity contentions rather than relying on any assumed estoppel.

If you are evaluating the patent as prior art (the more likely real-world use):

• US 4,718,094 (Bahl et al., filed 1986-03-27, issued 1988-01-05, priority 1984-11-19) is squarely § 102(b)/§ 102(a) prior art against modern speech-recognition patents. Its disclosure is unusually rich for that purpose: simplified phone machines with per-label replacement probability values, uniform length pseudo-distributions, tree-structured baseforms with shared-prefix simultaneous matching, stack decoding with a likelihood envelope and Δ-good/Δ-bad path pruning, and the label polling vote/penalty scoring tables. Notably, its § 371 priority chain includes a continuation-in-part of Ser. No. 672,974 (1984-11-19) and a further CIP of Ser. No. 738,930 (1985-05-29) — check which claims get which priority date before asserting it, because the 1984 and 1985 subject matter may not be co-extensive. Petitioners have already found this patent useful for exactly this: Amazon put it in the record in IPR2023-01194.
• Note the Board's claim-construction methodology in the related proceedings if you are citing the '094 patent alongside it: in Apple v. Zentian, the Board gave "feature vector" its broadest supportable reading consistent with the claim's "comprising a plurality of extracted and/or derived quantities," rejecting a narrow "pure" construction. That is a helpful signal on how the Board reads speech-recognition claim language generally.

Bottom line: there are no proceedings to brief. The correct output of this task is a negative finding plus the two false positives flagged above. If your client's exposure question is real, it is a patent-term-expiration question, not a PTAB question — route it to whoever on the team can pull the PatentCenter maintenance-fee and term-extension records, because that is the only document that will end the conversation.

I'll research the assignment record for US 4,718,094. Let me run targeted searches.

I found a significant detail — let me dig further into the IPR and the assignment records.

Let me try to pin down the recorded assignment details and check for any post-issuance transfers.

I've hit my search-step limit. Here's the ownership-chain reconstruction based on what the sources actually support — and I flag explicitly where the record is thin rather than filling gaps with inference.

---

Inventors

All four named inventors were IBM employees at the time of filing, based in Westchester/Dutchess County, New York — consistent with IBM's Thomas J. Watson Research Center (Yorktown Heights) and its speech-recognition group:

Source: the printed patent front page reproduced as Exhibit 1052 in [[Amazon.com Services](/litigations/by-defendant/Amazon.com%20Services) LLC](/litigations/by-plaintiff/Amazon.com%20Services%20LLC) v. Zentian Ltd., IPR2023‑01194 (PTAB, filed 2023‑07‑12), which reproduces the '094 cover page (inventors, assignee "International Business Machines Corp., Armonk, N.Y."). See https://www.docketalarm.com/cases/PTAB/IPR2023-01194/Amazon.com_Services_LLC/07-12-2023-Petitioner/Exhibit-1052-Exhibit_1052_US_Patent_No_4,718,094_to_Bahl/

Unusual patterns: None. This is the classic single-assignee corporate research-team pattern: four co-workers, one assignee, no individual inventor retaining rights. No evidence that any inventor departed IBM within 12 months of filing, and no inventor appears as an assignor in any post-issuance transfer. The "departure precedes a fire-sale" tell is not present.

---

Original assignee

International Business Machines Corporation (Armonk, NY) — named as assignee on the issued patent (Appl. No. 06/845,155; filed 1986‑03‑27; granted 1988‑01‑05).

• Primary line of business: mainframe/computing hardware and software; IBM was, at the relevant time, the dominant vendor of large-scale data-processing systems.
• Did IBM ship a product embodying the claims? The patent itself is a research/algorithmic disclosure (statistical speech recognition: fast acoustic match, detailed match, polling, stack decoding, auditory-model acoustic processor). IBM did commercialize dictation/speech products later (e.g., its continuous-speech dictation line), but I cannot confirm from the sources consulted that any shipped IBM product embodied these specific claims. I am not going to assert it.
• Current status: Operating. IBM has never been in bankruptcy, never dissolved, and never filed Chapter 7/11. It remains an active, publicly traded operating company (NYSE: IBM). The patent itself expired 2005‑01‑05 (anticipated expiration per Google Patents legal status), i.e., it lapsed at end of statutory term and is now public domain.

---

Assignment timeline

Important limitation, stated up front: The Google Patents legal-events record (the authoritative text supplied) shows exactly one recorded assignment — the original inventor → IBM assignment — and no post-issuance assignment of any kind. I was unable to retrieve the specific reel/frame number for that single assignment from the sources I could reach before exhausting my search steps. I am therefore not inventing a reel/frame. What the record supports is:

• 1986‑03‑27 (execution date coincides with the filing/priority event on Google Patents; exact execution date not separately confirmed) / recorded on or about 1986‑03‑27
  • Reel/Frame: not retrieved — verify at USPTO Assignment Center (search "4718094"). Do not treat the reel/frame as established until confirmed.
  • Conveyance: Assignment (ASSIGNMENT OF ASSIGNORS INTEREST — the standard inventor-to-corporate-employer assignment).
  • Assignor: Steven V. DeGennaro; Robert L. Mercer; Lalit R. Bahl; Peter V. deSouza (individually).
  • Assignee: INTERNATIONAL BUSINESS MACHINES CORPORATION, a corporation of New York.
  • Correspondent: Not retrieved from the sources consulted. Note that the prosecution attorney of record on the '094 cover page is listed only as "Marc A. ___" (surname truncated in the reproduced front page) — that is the prosecuting attorney, not an assignment correspondent, and should not be conflated with one.
  • Context: Internal corporate employment assignment — inventors assigned their rights to their employer at filing. Not a fire-sale, securitization, or transfer-to-asserter.

There are no subsequent recorded assignments. Per the rule for this task: no post-issuance assignment is itself a finding — it means the original assignee (IBM) retained ownership of US 4,718,094 for the entire life of the patent, through expiration on 2005‑01‑05.

---

Timeline diagram

timeline
    title Ownership of US 4718094
    1984 : Priority date
    1986 : Application filed
         : Inventors assign to IBM
    1988 : Patent issued
    2005 : Patent expired

---

NPE / troll-pattern signals

1. Shell-entity transfer — Not present. No assignment to any entity with an "IP / Patents / Licensing / Holdings / Ventures" suffix; no LLC assignee appears anywhere in the record. The only assignee in the chain is IBM itself.

2. Known asserter in the chain — Not present. IBM does not appear on any public NPE list (Acacia, Marathon, IV, IPNav, Wi‑LAN, Converso/Mosaid, Vringo, Pendrell, Innovatio, MPHJ, Lumen View, Round Rock, Erich Spangenberg entities), and no such entity appears in the '094 record.

3. Repeat correspondent across the chain — Unclear / no finding. A repeat-correspondent signal requires two or more recordings with the same attorney of record. Here there is only one recorded link, and I could not retrieve its correspondent. One link cannot establish recurrence. Not a finding.

4. Cascading transfers — Not present. Zero consecutive assignments through chained LLCs; no transfers within 24 months or at all after 1986.

5. Pre-litigation transfer — Not present. The '094 was never the subject of an infringement suit and the chain has no post-1986 transfer, so there is no "assignment within 6 months before suit" to assess. (The '094 does appear as prior art, not as an asserted patent, in IPR2023‑01194 — Amazon v. Zentian — and in the related Zentian v. Apple appeals; those are citations to the '094, not assertions of it.)

6. Bankruptcy fire-sale — Not present. IBM has never filed for bankruptcy protection; no portfolio sale of this patent occurred.

7. Privateering — Not present. No operating-company-to-NPE transfer exists in the chain; IBM retained the patent and never farmed it out to an NPE.

8. Defensive aggregator (anti-NPE) — Not present as a chain event. The chain does not terminate at RPX, AST, LOT, Unified Patents, or OIN. IBM simply kept the patent until it expired on its own.

---

Verdict

Insufficient data — in the benign sense defined by this task: only the original assignment exists.

The entire recorded chain is a single inventor → IBM corporate employment assignment (dated on/around 1986‑03‑27, on the same date the application was filed), with no subsequent recorded conveyance before the patent's anticipated expiration on 2005‑01‑05. That pattern — an operating-company original assignee retaining title for the patent's full life and never transferring to a shell LLC, licensing entity, or known asserter — produces zero strong and zero weak NPE signals. On the evidence available, this is a no-asserter, original-owner-retained outcome rather than any form of NPE activity. I cannot claim a stronger conclusion (e.g., "Operating-company assertion" or "Defensive/non-asserting") because IBM neither asserted the '094 nor placed it into a defensive aggregator; and I cannot give the exact reel/frame or correspondent with the sources I reached.

Verify at: USPTO Assignment Center — search patent number 4718094 (https://assignmentcenter.uspto.gov/ ; mirror at https://assignment.uspto.gov/patent/index.html) — to confirm (a) the single assignment's reel/frame, (b) its correspondent, and (c) the absence of any later conveyance. Cross-check the Google Patents legal-events tab (https://patents.google.com/patent/US4718094/en), which lists only the 1986-03-27 assignment and the 2005‑01‑05 expiration.

Contradiction flagged with the earlier litigation section: the prior "litigation summary" reported no litigation touching the '094 at all. The current search surfaced **IPR2023‑01194 (Amazon.com Services LLC v. Zentian Ltd.)**, where US 4,718,094 appears as Exhibit 1052 (prior art) against Zentian's U.S. Patent 10,062,377 — matching the earlier finding's characterization of the '094 as a cited reference rather than a patent-in-suit, but it is a second, distinct citation instance (the earlier section cited only the Apple v. Koss IDS and the KR 20150104615 citation). Neither contradicts the ownership conclusion.

I'll research the cited prior art for US 4,718,094 on the USPTO record and related databases.

I found the key source — the "References Cited" section. Let me gather the complete list of cited references and verify details.

I've located the authoritative front-page "References Cited" list for US 4,718,094. Note: this is the patent's own prior-art citation list — not litigation. Below is the enumerated analysis, with an explicit correction/refinement to the earlier Litigation Summary.

---

Prior-Art Analysis — US 4,718,094 ("Speech recognition system")

1. Patent identification (per the authoritative record)

Claim landscape (context for the §102 mapping). The exposed method claims are directed to an approximate "fast" acoustic match in which each phone machine is simplified by assigning a single replacement value to all label probabilities for a given label, plus corollaries on (i) the replacement value being ≥ the maximum actual label probability (claim 5), (ii) converting the length distribution to a uniform pseudo-distribution (claim 6), (iii) a minimum non-zero length floor (claim 7), (iv) end-time/start-time distribution propagation (claims 8, 14), (v) summing phone match values into a word score (claim 9), and (vi) a subsequent detailed match on the surviving candidate words (claim 11). One rendering (uspto.report) also shows dependent claims numbered in the 60s referencing claims 5–8, indicating the patent carries an extensive claim set beyond the first 15 claims. Flag: the full claim count and the polling/score-combination claims should be confirmed against the printed patent before relying on specific claim numbers.

---

2. Source and completeness caveat

The enumerated references below are taken from the patent's own front-page "(56) References Cited" list as reproduced in the record (Google Patents; the free full-text copy at patentimages; and Petitioner's Exhibit 1052 in IPR2023-01194, which is a copy of this exact patent). The U.S. patent list is reproduced essentially in full. The "Other Publications" list below is partially truncated in the accessible sources (a few entries cut off mid-line, e.g., "J. F. Mari et al. ..."). The "Foreign Patent Documents" subsection was not surfaced in the sources I retrieved — I cannot confirm whether the printed patent contains foreign citations. Treat this as a near-complete but not certified-exhaustive enumeration.

---

3. U.S. Patent Documents cited in US 4,718,094

All are classified in speech-analysis/recognition subclasses (179/1 SA, 179/1 SB, 179/1 SD, 381/43, 381/.43, 364/513.5), confirming the art field. Titles/assignees are noted only where I am reasonably confident; where I am not, the entry is marked [verify] rather than guessed.

Key observation: references #16–#18 carry issue dates after the '094 filing date (1986-03-27). They cannot be §102(a)/(b) art on their face; if they were considered at all they would have to be §102(e) art measured by their own U.S. filing dates. This is consistent with them being examiner/supplemental cites during the 1986–1987 prosecution.

---

4. Foreign Patent Documents

Not surfaced in the sources retrieved. Unable to confirm whether any foreign citations appear on the printed patent. Flag for verification against the printed front page.

---

5. "Other Publications" cited (non-patent literature)

---

6. Bottom line — most relevant prior art and the §102 position

Ranked relevance for US 4,718,094:

1. Bahl et al., "Faster Acoustic Match Computation," IBM TDB vol. 23(4), Sep. 1980 (O1) — the single most on-point reference, squarely about speeding up the acoustic match and originating from the same IBM group. This is the best §102 candidate for the broadest match-acceleration concept.
2. Baker et al., US 4,348,553 (Sep. 1982) and Welch et al., US 4,319,085 (Mar. 1982) — closest U.S. patents on HMM/probabilistic-decoder speech recognition; §103/background, not anticipatory.
3. Bahl, Jelinek & Mercer, "A Maximum Likelihood Approach to Continuous Speech Recognition," PAMI-5(2), Mar. 1983 (O2) — the foundational detailed-match paper, cited by the applicant inside the specification itself (§103 art / admitted background).
4. Sakoe et al., US 4,158,750 (Jun. 1979) — DTW template-matching background.

§102 conclusion (preliminary and hedged): On the face of the record, the cited prior art does not appear to anticipate claims 1, 5, 6, 7, 8, 9, 11, or 14 under §102. The record's own labeling (a long list of similar-field speech patents plus 1985–1986 conference literature) is characteristic of a §103/background citation set, not an anticipation set. The apparent novelty resides in the specific approximations — (a) replacing all transition-specific label probabilities for a label with a single value (claim 1), (b) that value being at least the maximum actual probability (claim 5), and (c) the uniform length pseudo-distribution with a minimum-length floor (claims 6–7) — none of which the identified references disclose.

Critical caveat: I have not retrieved and read the full text of each cited reference. The §102 reach stated above is a triage screen based on citation metadata, fields of art, and dates — it is not an element-by-element anticipation opinion. A true §102 analysis requires the full disclosures of at least O1, O2, US 4,348,553, US 4,319,085, and the post-filing references #16–#18 (for §102(e) purposes). I am explicitly flagging items [verify]/[unverified] rather than supplying titles I cannot confirm.

---

7. Correction/refinement to the prior Litigation Summary

The earlier section reported (correctly) that no assertion litigation involving the '094 patent was found, and flagged IPR2022-00053 (Apple v. Koss) as an IDS citation only. In this pass I found a second, more substantive appearance of the '094 patent in a PTAB proceeding — still not an assertion of '094, but it should be recorded:

• IPR2023-01194, [[Amazon.com Services](/litigations/by-defendant/Amazon.com%20Services) LLC](/litigations/by-plaintiff/Amazon.com%20Services%20LLC) (and Amazon Web Services, Inc.) v. Zentian Limited, P.T.A.B., filed 2023-07-12, instituted 2024-01-10, FWD 2024-06-11, status "Final Written Decision – Appealed."
  • Patent at issue: US 10,062,377 (Zentian Ltd.; Respondent Application No. 14/788,164).
  • US 4,718,094 = Petitioner's Exhibit 1052, i.e., it was relied upon as prior art against the '377 patent.
  • Sources: https://ipverse.greyb.com/ptab-web/cases/case-details/IPR2023-01194 ; https://www.docketalarm.com/cases/PTAB/IPR2023-01194/Amazon.com_Services_LLC/07-12-2023-Petitioner/Exhibit-1052-Exhibit_1052_US_Patent_No_4,718,094_to_Bahl/
  • Related district-court matters listed in Zentian's mandatory notices: Zentian Ltd. v. Apple Inc., 3:23-cv-02921 (N.D. Cal.) and Zentian Ltd. v. Amazon.com, Inc., 6:22-cv-00123 (W.D. Tex.). Neither is a case in which '094 was asserted — '094 is cited as an invalidating reference against a different patent.

This refines but does not contradict the prior Litigation Summary: '094's only litigation/proceeding roles are as a cited prior-art reference, never as a patent-in-suit.

Forward citations for completeness (not prior art): US 4,819,271, US 4,833,712 (IBM sibling applications), KR 20150104615 A, DE 3337353 C2 — all cite '094; they post-date it and are irrelevant to its §102 analysis.

I'll pull the claim text and the reference/citation record I need, since the authoritative fetch above contains the description but cuts off before the claims and does not include a discrete "Prior Art"/"References Cited" list.

Obviousness Analysis — U.S. Patent No. 4,718,094 ("Speech recognition system," Bahl et al., IBM)

0. Scope, sources, and two required flags

Flag A — the "Prior Art section" is not in the authoritative fetch. The authoritative Google Patents text provided to me contains a "Definitions" block, a "Prior art keywords" tag list (word, label, phone, labels, given), and a "DESCRIPTION OF PRIOR AND CONTEMPORANEOUS ART" section written by the applicants. It does not contain the front‑page "References Cited" list. To obtain that list I used secondary mirrors of the printed front page (uspto.report/patent/grant/4718094 and the PTAB exhibit copy of the patent at Docket Alarm, IPR2023‑01194 Ex. 1052). The reference list below is therefore secondary‑sourced; where that list or those sources conflict with the authoritative fetch, I follow the authoritative fetch. There is no direct conflict — only absence.

**Flag B — the applicant's own "Prior and Contemporaneous Art" narrative is heavily admission‑laden, and those admissions are the single most powerful §103 material in the document. Examples from the authoritative text:

• "Acoustic matching is moderately effective. However, acoustic matching normally requires considerable computation. Hence, a computationally inexpensive methodology which could be used instead of, or in conjunction with, the acoustic matching is a desired end."
• "A key question in linguistic decoding has therefore been how to determine Pr(Y|W) … without requiring inordinate computation time and without sacrificing accuracy."
• "One known technique for storing the tree is referred to as 'preordering' in the prior technology."
• "Two prior art approaches for finding the most likely word sequence are Viterbi decoding and single stack decoding." (both described in Bahl/Jelinek/Mercer 1983, §§V and VI)
• "Language models have been reported in the literature."

These are statements of the problem and of what was known, and they supply the motivation element of the obviousness case almost verbatim.

Cross‑reference update to the litigation section. Consistent with the earlier negative finding, the '094 patent is not a patent‑in‑suit. But the search surfaced a concrete, citable instance of its use as prior art: IPR2023‑01194, [[Amazon.com Services](/litigations/by-defendant/Amazon.com%20Services) LLC](/litigations/by-plaintiff/Amazon.com%20Services%20LLC), Petitioner's Exhibit 1052 = U.S. Patent No. 4,718,094 to Bahl, filed 2023‑07‑12 (docketalarm.com, keyed to patent 10,062,377). That refines — it does not contradict — the earlier conclusion: the '094 patent appears in PTAB proceedings as a printed publication prior‑art exhibit, not as an asserted patent.

---

1. The claims being tested (97 claims, 13 independent)

The claim set resolves into three families, which map exactly onto the three "aspects" the applicants themselves identify in the specification:

I will not re‑describe this content (it was covered in the earlier sections); I use it only as the claim‑taxonomy scaffold.

---

2. Level of ordinary skill in the art (POSITA), as of 1984‑11‑19

A POSITA would have: an M.S. or Ph.D. in EE, CS, or statistics (or a B.S. plus 3–5 years in speech recognition); working command of hidden Markov modeling as of Bahl/Jelinek/Mercer 1983, including the forward‑backward/Baum‑Welch estimation algorithm and Viterbi decoding; the vector‑quantization "label/feneme" front end; and large‑vocabulary decoder architecture (fast‑match → language model → detailed‑match → search), which was the standard IBM/DRAGON/HARPY pipeline by 1983. This is the level reflected by the patent itself (PL/I on an IBM MVS system, APAL on FPS‑190L array processors).

---

3. Prior art available against the 1984‑11‑19 priority date

Important threshold point that a rigorous analysis must make: many references on the printed front page cannot be §103 art against this priority date. The front page mixes in articles and U.S. patents from 1984–1987 that post‑date the Nov. 19, 1984 CIP priority date. Unless a U.S. patent's filing date predates the invention date (§102(e)), these are not available. I have partitioned accordingly.

Group 1 — Printed publications, §102(b) art (published before 1983‑11‑19), fully available:

Group 2 — U.S. patents whose issue dates fall in 1984–1986. Citable only on their filing dates (§102(e)); verification required: US 4,467,437 (Tsuruta, 8/1984 — issued before 11/19/1984, so it is also §102(a) art); US 4,481,593 (Bahler, 11/1984); US 4,489,434 & 4,489,435 (Moshier, 12/1984); US 4,513,436 (Nose, 4/1985); US 4,516,215 (Hakaridani, 5/1985); US 4,559,604 (Ichikawa, 12/1985); US 4,587,670 (Levinson, 5/1986); US 4,593,367 (Slack, 1/1986); US 4,618,983 (Nishioka, 10/1986).

Group 3 — References that are not available against 1984‑11‑19 and should be excluded from any ground: Cravero et al. (CSELT, Apr. 1986); "Composite Fenemic Phones," Research Disclosure No. 256 (Aug. 1985); Rabiner et al. (Jul.–Aug. 1985); Bourlard et al. (1985); Paul (1985); Juang et al. (1985); Kamp (Oct. 1985); Sugawara (1985); Schwartz et al. (1984/1985); Mari et al. (1985); and the inventors' own "Improving the Training…" and stack‑decoder copending applications (filed 1985; commonly owned, and thus also disqualified under §103(c)). Their presence in the file wrapper reflects prosecution‑era citation practice, not §103 reachability.

Supplementary art I located (not on the printed front page), offered with a confidence caveat: Ferguson, "Variable duration models for speech," Proc. Symp. on Applications of Hidden Markov Models to Text and Speech, IDA‑CRD, Princeton, pp. 143‑179 (1980). Multiple independent citation records confirm this 1980 date and venue. It is the canonical treatment of duration/dwell modeling in HMMs, including fixed‑duration and uniform‑duration state models. This is §102(b) art and is directly on point for the uniform‑length‑distribution claims. (Confidence: high on existence/date; I have not read the paper, so I do not assert its precise equations.)

---

4. Ground‑by‑ground obviousness

Ground 1 — Family A, base claim 1 (and claims 2–13, 14–17, 18–26, 27–40)

Proposed combination: Bahl/Jelinek/Mercer 1983 (primary) + Jelinek 1976 + Bahl et al., "Faster Acoustic Match Computation," IBM TDB (Sept. 1980) + Hart/Nilsson/Raphael, "A formal basis for the heuristic determination of minimum cost paths," IEEE Trans. SSC‑4, pp. 100‑107 (1968).

Element mapping.

• Claim 1's preamble — "at least one sequence of phones … phone machine … actual label output probabilities, each actual label probability representing the probability that a subject label is generated at a given transition" — is verbatim Bahl 1983 (and the patent concedes this: "the Markov model … has been described in various articles such as 'A Maximum Likelihood Approach…'"). Indeed the patent's own FIG. 3 / tr1–tr13 / p[1]…p[200] structure is the Bahl 1983 model.
• The sole inventive step in claim 1 is: "replacing by a single specific value the actual label probabilities for a given label at all transitions at which the given label may be generated in a particular phone machine," then "determining the probability of a phone generating the labels in the string based on the simplified phone machine."
• The specification itself characterizes the objective as computational: "the number of computations is high," and the replacement "reduces greatly" the required computations. The Sept. 1980 IBM TDB is titled Faster Acoustic Match Computation.

Motivation / rationale (KSR‑compliant).

1. Known problem, known fix: The exact match of Bahl 1983 is O(states × transitions × alphabet) per phone per frame. Collapsing one dimension (transitions) is the most direct available reduction and yields a strictly smaller computation for the same lattice.
2. Admissible‑bound reasoning was a known, general technique. Hart/Nilsson/Raphael 1968 (itself cited as reference 22 in Jelinek 1976, i.e., squarely in the same literature the POSITA was reading) teaches choosing an approximation guaranteed to be optimistic so that pruning remains safe. Claim 5 / 19 / 21 / 60 (p′(yₖ) ≥ max p(yₖ)) is precisely an admissible upper bound, and the patent concedes exactly this rationale: "since the replacement value is preferably an overestimation, the resulting match score is not less than … without the replacement."
3. Predictable result: Using the maximum produces an overestimate, a known and predictable consequence; the patent claims no unexpected accuracy benefit from the max‑value choice itself.
4. Design incentive: large vocabulary (~5,000 words) and real‑time decoding — the market pressure the applicants themselves recite in "Prior and Contemporaneous Art Pertaining to Total Scores."

Dependent claims.

• Claim 2 / 16 / 59 — tree of baseforms sharing common branches. This is the weakest point of the patent's case: the specification admits "One known technique for storing the tree is referred to as 'preordering' in the prior technology," and Baker's DRAGON (1975 paper; US 4,348,553, Sept. 1982) discloses an integrated Markov network in which shared prefixes are traversed once, with outputs produced by state traversal, and Baker expressly states the total number of computations is linear in utterance length because "alternate paths are recombined." Motivation: eliminating duplicated prefix computation when processing all words against one label string.
• Claims 3–4 — cap the match at a predefined maximum number of labels (~100 centiseconds). Motivation is express in the specification ("decoding delay is reduced" and "problems in comparing the scores of short words with long words are substantially avoided") and the technique — truncating a lattice search at a fixed horizon / beam — is standard in Lowerre‑type and Jelinek‑1976 decoders.
• Claims 5, 13, 19, 21, 60 — the max‑value choice. See rationale 2 above.
• Claims 6, 7, 12, 20, 22, 28, 61, 62 — uniform length pseudo‑distribution, with a minimum‑length floor. Combine with Ferguson 1980 (duration/dwell modeling in HMMs, including uniform/fixed duration state models). Motivation: (i) it makes the recursion constant‑time per frame — the patent itself quantifies the gain, "forty multiplications and twenty additions" → "one multiplication and one addition"; (ii) it is the standard engineering trade of duration fidelity for computation; (iii) taking L_min as the smallest length with nonzero probability is the natural truncation of a pseudo‑distribution.
• Claims 8–11, 14–17, 23–26, 29–40 — end‑time/start‑time distribution recursion, log‑sum match value, summing per‑phone match values over a baseform, then detailed match on survivors. Every one of these is disclosed or rendered obvious by Bahl 1983 (end‑time distribution; summing end‑time probabilities; log of the sum; start‑time distribution of the next phone derived from the previous phone's end‑time distribution). Claim 11's "perform a detailed acoustic match with an unsimplified phone machine for each word of the selected number" is the two‑pass coarse‑then‑fine architecture that the applicants admit was the reason for the fast match in the first place.
• Claims 25, 29, 39 — the θₙ recursion with L_min = 0. This is the algebraic consequence of factoring a constant out of the Φₙ sums (the specification demonstrates exactly this derivation). Where the result is "the necessary consequence of the combination," it is obvious as a matter of law‑like reasoning under In re Kao / MPEP 2144.02.
• Claim 31 (first J+K labels, where K = number of states). The "+K" is dictated by the need to let all states of the widest machine finish a transition; a POSITA would select K as the maximum state count from the same tree/statistics the machine already stores. Routine optimization.

Conclusion on Ground 1: Claims 1, 5–12, 14–16, 19–22, 26–34, 36–39 are obvious over Bahl 1983 + Jelinek 1976 + IBM TDB 9/1980 + (Ferguson 1980 for the duration claims). Claims 2/16/59 are obvious further in view of Baker/DRAGON and the applicants' own "preordering" admission. Claims 3–4 obvious further in view of the admitted decoding‑delay problem and standard pruning.

---

Ground 2 — Family B: polling (independent claims 41, 51, 64, 69, 97; dependents 42–63, 65–68, 70)

Proposed combination: Bahl/Jelinek/Mercer 1983 + Jelinek 1976 + Bahl et al., "Interpolation of Estimators Derived from Sparse Data," IBM TDB (Sept. 1981) + Bahl et al., "Faster Acoustic Match Computation," IBM TDB (Sept. 1980).

Element mapping.

• Claim 41(b) "determining label votes, each label vote representing the likelihood that a respective label is produced when a given word is uttered" and 41(c)/(42) accumulating/summing votes = computing, for each word, a label‑marginal score from the same trained label‑output and transition statistics Bahl 1983 already produces for the detailed match. Claim 50's step (t)–(v) ("determining transition likelihood counts and label likelihood counts from training data … producing an expected label distribution for each word … each label vote being derived from the expected label distribution") makes the provenance explicit — the votes are literally derived from the Bahl 1983 bᵢⱼₖ statistics.
• Claim 52's "logarithmic probability … being the label vote" is the standard log‑probability bookkeeping (no multiplication in the inner loop), universally used in the art.
• Claims 46, 53, 65, 67 — the penalty table, "the likelihood of the given label not being produced according to the model," summed over labels not in the string, with votes summed over labels in the string: this is the complementary ("null") probability normalization that any probabilistic model requires so the scores are comparable; the sum of "vote + penalty" over the full alphabet is the normalization constant.
• Claims 44, 54, 56 — dividing by the number of labels considered, and the "highest relative likelihood score at each successive interval" tie‑break.

Motivation / rationale.

1. The applicants state the motivation outright: "acoustic matching normally requires considerable computation. Hence, a computationally inexpensive methodology which could be used instead of, or in conjunction with, the acoustic matching is a desired end." A desired end recited by the applicant is a §103 motivation.
2. The technique itself — precompute, for each word, an expected per‑label distribution, then score an incoming label string by table lookup — is a straightforward marginalization of the exact HMM score, i.e., an obvious simplification of Bahl 1983 over the label dimension. Where the art discloses alternatives and a POSITA selects the computationally cheaper with a predictable (increased) error rate, that is KSR "obvious to try."
3. The sparse‑data interpolation TDB (Sept. 1981) supplies the estimation machinery for producing stable expected distributions/penalties from limited training counts, so the vote and penalty tables are enabled and motivated.
4. Length scaling (÷ number of labels) addresses the admitted problem of "comparing match scores for words of different lengths" — the same problem the patent says its J‑label truncation addresses in Family A. The two families solve one problem two ways; using both is a routine design choice, and the patent itself contemplates it: "the acoustic fast match and the polling may be used in conjunction."

Conclusion on Ground 2: Claims 41–70 and 97 are obvious over Bahl 1983 + Jelinek 1976 + "Faster Acoustic Match Computation" (1980) + "Interpolation of Estimators" (1981). The strongest dependent‑claim candidates for surviving a challenge are claim 56 (the specific "highest relative likelihood score at each successive interval" conservative‑selection rule) and claims 48/49 (limiting votes/penalties to labels before the most‑likely end time, and repeating at successive intervals) — these are more narrowly tied to a specific decoding‑timing strategy than to the generic table‑lookup idea.

---

Ground 3 — Family C: combining plural independent word scores (independent claims 71, 84, 92; dependents 72–83, 85–96)

Proposed combination: Jelinek/Bahl/Mercer 1975 ("Design of a Linguistic Statistical Decoder") + Bahl/Jelinek/Mercer 1983 + Baker, "The DRAGON System — An Overview" (1975) / US 4,348,553 (Sept. 1982).

Element mapping.

• Claim 71: "generating a first word score … based on an acoustic match first algorithm; … a second word score based on a second independent algorithm which differs from the first algorithm; and forming a total word score … from at least the first word score and the second word score."
• Claim 84/92: N word scores from N independent algorithms, combined.
• Claim 72/82/95: weighting at least one score before combining.

The combination is squarely taught. The decoded quantity in both Jelinek/Bahl/Mercer 1975 and Bahl 1983 is Pr(W)·Pr(Y|W)/Pr(Y) — i.e., the product (equivalently, the weighted sum in the log domain) of two independently derived scores: a language‑model score Pr(W) and an acoustic‑model score Pr(Y|W). Claim 72's "weighting at least one of the word scores relative to the other word scores and combining the word scores after weighting" is the interpolation/exponent‑weighting already standard for combining a language model with an acoustic model. Going from two scores to three (fast match + detailed match + polling, or polling + acoustic) is the KSR "two known elements → three" case: the patent itself says "the total word score … may be a combination of more than two independent word scores," which concedes the generalization.

Baker/DRAGON supplies the more aggressive form — multiple knowledge sources (acoustic‑phonetic, lexical, phonological, syntactic‑semantic) each represented as a probabilistic function of a Markov process and folded into one integrated Markoov network to compute a single posterior, expressly so the system "function[s] despite high error rates from individual knowledge sources." That is the exact rationale the patent gives for combining scores ("The use of the total scores — rather than word scores from a single algorithm — results in enhanced accuracy with no significant increase in computational requirements").

Motivation / rationale.

1. Error decorrelation: combining classifiers/scores derived from different statistics tends to reduce error — the same insight Baker states for DRAGON's knowledge sources, and the same insight already embodied in the acoustic‑plus‑language‑model combination of Bahl 1983.
2. No added complexity: in log space, combining scores is an addition; the patent says so ("with no significant increase in computational requirements").
3. Reuse: claims 79–83 (fast‑match score reused as both a candidate filter and a total‑score component) are an efficiency‑motivated reuse of already‑computed numbers — the strongest possible case under KSR ("a person of ordinary skill is a person of ordinary creativity"; using a computed quantity for its evident second purpose).
4. Claims 80–81 (different sets of phone machines, one fenemic and one phonetic). Motivation: Bahl 1983's phonetic‑phone model is precise but parameter‑heavy; a fenemic/label‑level model is parameter‑light. Running both and combining is the obvious way to capture both granularities. The patent's own data (30–50% error reduction) is asserted but not tied to a comparison against the closest prior art — see §6.

Conclusion on Ground 3: Claims 71–96 are obvious over Jelinek/Bahl/Mercer 1975 + Bahl 1983 + Baker 1975/US 4,348,553. The ordering of the claims does not change the analysis; note that claim 78 is drafted with a dependency irregularity (it recites "The method of claim 76" while reciting subject matter that corresponds to claim 77's "polled word score … being/corresponding to the second word score"), which affects the scope of the challenge to claim 78 but not its substance.

---

5. Consolidated rejection chart (representative independent claims)

---

6. Best available non‑obviousness arguments (and why they are weak here)

a. "The 30–50% error‑rate improvement shows unexpected results." The specification states, "the use of a total word score … results in a 30% to 50% error rate improvement over a speech recognition system which simply applies a detailed match to candidate words derived by a fast match algorithm." This fails as a secondary‑consideration showing because (i) it compares the applicants' own two configurations, not the claimed invention against the closest prior art; (ii) there is no nexus statement, no statistical significance, and no test‑corpus identification; and (iii) an improvement attributable to the known benefit of combining independent scores is not "unexpected" under KSR.

b. "The specific choice of the maximum as the replacement value is critical." Claims 5/19/21/60 limit to ≥ max. This is the only choice that guarantees the required overestimation, so the claim is coextensive with the stated function — obvious as a design choice dictated by the objective (MPEP 2144.04).

c. "Uniform length distribution with a min‑length floor is a deliberate simplification with retained accuracy." This is the strongest argument, but it is undercut by Ferguson 1980 and by the specification's own arithmetic showing the simplification is purely a computational trade.

d. Narrow claims with best survival odds: claim 56 (successive‑interval "highest relative likelihood" assignment), claims 48–49 (limiting votes/penalties to labels preceding the determined end‑time label; re‑combining at successive intervals), claim 31 (the specific J+K column‑score replacement using the maximum state count), and possibly claim 25/29/39 (the exact θₙ recursion form) — these recite specific algorithmic choices with no squarely on‑point prior‑art hit in Group 1. They are nonetheless vulnerable under In re Kao-type reasoning if shown to be the algebraic consequence of the combination.

---

7. Bottom line

• Independent claims 1, 14, 18, 27, 40, 41, 51, 64, 69, 71, 84, 92, and 97 are, in my assessment, obvious under §103 over the Group 1 art — principally Bahl/Jelinek/Mercer 1983 (PAMI‑5, no. 2, pp. 179‑190), Jelinek 1976 (Proc. IEEE 64:532‑556), and Bahl et al., "Faster Acoustic Match Computation," IBM TDB 23(4):1718‑1719 (Sept. 1980) — and, for Family C, Jelinek/Bahl/Mercer 1975 plus Baker/DRAGON (1975; US 4,348,553). The motivation element is unusually well‑supported because the applicants themselves state the problem ("without requiring inordinate computation time"), the desired end ("a computationally inexpensive methodology … is a desired end"), and one piece of prior art ("preordering").
• The duration/length claims (6, 7, 12, 20, 22, 28, 61, 62) additionally depend on Ferguson 1980, which I have confirmed by multiple independent citation records but have not read in full.
• Every reference on the front page dated 1985 or later must be excluded from a §103 case against the 1984‑11‑19 priority date, and the 1984–1986 U.S. patents must be checked for §102(e) filing dates before use. This materially shrinks the usable art from what the printed page suggests.
• The patent expired 2005‑01‑05 and, per the earlier finding (updated here), appears in PTAB proceedings only as prior‑art Exhibit 1052 in IPR2023‑01194 (Amazon.com Services LLC), not as a patent‑in‑suit.

---

8. Caveats and verification steps

1. Claims provenance. The 97 claims and the "References Cited" list were taken from secondary mirrors (uspto.report; Docket Alarm exhibit copy of the printed patent). The authoritative Google Patents fetch supplied earlier is truncated before the claims. Verify the claim text and the front‑page reference dates against the granted patent PDF (patentimages US4718094.pdf), particularly the Scarr year/volume anomaly and the exact wording of claim 78's dependency.
2. "Prior Art section." The Google Patents page as fetched does not expose a discrete "Prior Art" tab (only "Prior art keywords" tags); if the task assumed a populated Prior Art panel, that panel was not in the supplied text. This analysis therefore uses the applicants' own "DESCRIPTION OF PRIOR AND CONTEMPORANEOUS ART" plus the printed "References Cited."
3. Unread references. I have not read the full text of Ferguson 1980, Scarr, Haton 1982, Choy 1982, Das 1980, or the U.S. patents in Group 2. Their treatment above is based on titles, dates, and standard art knowledge, and should be confirmed before any position is taken.
4. Related‑application inconsistency (flagged, not corrected). A secondary source (a 2024 conference PDF reproducing the front page) states the '094 application is a CIP of Ser. No. 672,974 (Nov. 19, 1984), "which is a continuation‑in‑part of Ser. No. 738,930, May 29, 1985." That chronology is impossible (a Nov. 1984 case cannot be a CIP of a May 1985 case) and conflicts with uspto.report, which lists 672,974 and 738,930 as parallel CIP parents. Separately, the specification cites the stack‑decoder application as Ser. No. 06/738,911 (May 29, 1985) and the multi‑utterance baseform application as Ser. No. 06/738,933 (May 29, 1985), neither of which appears in the front‑page related‑application list. The authoritative fetch should govern; the discrepancy is worth resolving before relying on any priority/family argument.
5. If a formal validity opinion is needed, next steps would be: full‑text retrieval of Bahl 1983, Jelinek 1976, the two IBM TDBs, Baker 1975/DRAGON, and Ferguson 1980; a §102(e) filing‑date check on the Group 2 patents; and a §112 review of the several apparent claim‑drafting irregularities (e.g., "three structure" in claim 2; claim 78's dependency; the undefined antecedent for "θₙ" in claims 25/29/39, which recite formulas that do not appear in the text supplied).