US 5995928

I'll search for authoritative information on this specific patent number, including any CAFC 2026 docket activity.

Let me search more specifically for any CAFC or litigation activity involving this patent.

US Patent 5,995,928 — Summary

Important sourcing caveat first: I ran targeted searches for this patent on USPTO/CAFC-related materials. I did not find any CAFC 2026 docket, appeal, IPR, or district-court activity involving US 5,995,928. I also could not open USPTO PatentCenter/CAFC docket pages directly in these searches. So I can report the bibliographic and claim content (well-corroborated across several independent patent databases), but I cannot confirm any 2026 litigation or appellate activity — and absent evidence, you should treat "no CAFC 2026 activity found" as a negative search result, not a certified fact. Note also that this is a 1996-filed patent that reached its 20-year term, which makes active litigation unlikely.

Bibliographic data (v. US 5,995,928 A)

Abstract (as issued)

A speech recognition system for recognizing a word or words based on a user's continuous spelling. A speech recognition engine runs its decoder in forward mode so it continuously outputs an updated string of hypothesized letters as the user speaks. A spelling engine compares each hypothesized string to a vocabulary list and returns a best match. An optional early-identification unit presents the best-matching word to the user, possibly before the user finishes spelling.

Independent claims — plain language

Claim 1 (system). A speech recognition system that works on the word before it is fully spelled, as each letter is received. Two components: (a) a speech recognition engine that takes continuous spoken letters, forms letter hypotheses, and periodically emits an updated string of hypothesized letters representing a partial spelling; and (b) a spelling engine, connected to the recognizer, that has a vocabulary list and a confusability matrix (scoring how likely each hypothesized letter is to be confused with each letter of each vocabulary word). The spelling engine receives the updated partial strings and compares them to the vocabulary to pick the single best-matching word.

Claim 10 (method). The method counterpart: receive a user's continuously uttered letters; process them into a speech-recognizer-compatible signal as each letter arrives; analyze with a speech recognition engine to form letter hypotheses and periodically output an updated hypothesized string representing a partial spelling; and compare those updated strings to a preselected vocabulary using a spelling engine, repeating until a best match to a single vocabulary word is found.

Claim 14 (computer-implemented method). Similar to claim 10 but framed around an electronic speech signal and ending with an explicit output step: receive/process the uttered letters, analyze to form hypotheses before the word is complete, periodically output an updated hypothesized string representing the portion of the word, compare the string to preselected vocabulary words until a best match is found, and output the matched vocabulary word before the word has been completely spoken.

Plain-language gist of the whole patent

It solves a "fallback" problem: when a general speech recognizer fails on a name or unusual word, the user spells it. Rather than the older discrete-spelled approach (which prompts letter-by-letter and is slow), this patent recognizes continuous spelling. Because continuous letters blur together, the recognizer's hypotheses contain insertions, deletions, and misrecognitions — hence the spelling engine uses dynamic-programming matching plus a confusability matrix and a transition-cost table to score each vocabulary word, with a scoring module applying threshold and margin criteria to declare a match, enabling "early identification" before the spelling is finished. It reverts to a backward-pass (A*) decoder pass when the user stops speaking and no match was found.

Notable associated documents (from search results)

• A PCT search report (WO 2017084334 A1) cites US 5,995,928 A (SpeechWorks Int. Inc.), 30 Nov 1999 as prior art against claims 1–18.
• Google Patents "Cited By" lists ~113 later publications (e.g., "spelling fallback during large-scale speech recognition," "voice spelling in an audio-only interface").
• Non-patent literature cited includes Davis, Let Your Fingers Do the Spelling (1991) and MIT SUMMIT system papers.

Bottom line on the CAFC 2026 question

No evidence was found of any 2026 Federal Circuit docket, appeal, or pending litigation for 5,995,928 in the sources I could reach. Given the patent's expired status (term ended 2016-10-02), active appeal activity is improbable. If you need a definitive docket check, the authoritative sources would be USPTO PatentCenter (application 08/720,554), the USPTO assignment record, and the CAFC PACER/docket — none of which I could query directly here, so I flag this as a limitation rather than a confirmed absence.

I'll research litigation involving US patent 5995928. Let me search multiple sources.

Let me search litigation-specific sources.

Let me check the litigation portals and other specific sources.

Litigation Search Results — U.S. Patent No. 5,995,928

Bottom line

I found no evidence of any litigation (district court, ITC, or Federal Circuit appeal) involving U.S. Patent No. 5,995,928 specifically.

Based on the sources I was able to reach, U.S. 5,995,928 ("Method and apparatus for continuous spelling speech recognition with early identification," SpeechWorks International, Inc.; now shown as assigned to Nuance Communications, Inc.) does not appear in any patent-litigation docket record. The Unified Patents patent page for US‑5995928‑A (https://portal.unifiedpatents.com/patents/patent/US-5995928-A) lists the assignee and prior art but returns no litigation entries for this patent number. Likewise, the Google Patents page (https://patents.google.com/patent/[US5995928](/patent/US5995928)/en) shows only "Cited By" references and family/assignment data — no "Litigation" section.

Important caution — do not confuse this patent with a related case

My searches surfaced a well‑known SpeechWorks lawsuit, Elliott Davis v. SpeechWorks International, Inc. (W.D.N.Y., filed July 15, 2003; also a companion Nuance action), which produced a 2006 Markman/claim‑construction decision (Davis v. SpeechWorks Int'l, W.D.N.Y., Sept. 29, 2006).

That case is NOT about U.S. 5,995,928. The asserted patent there was U.S. Patent No. 5,231,231 (Davis's "independent error templates" patent), and the accused product was a Vodafone speech system using "decoy words." The '928 patent at issue here is a different patent, a different inventor set (Nguyen and Marx), and a different owner chain. I flag this because automated searches keyed on "SpeechWorks + spelling + patent" can easily conflate the two.

Similarly, "5995928" is visually similar to several other patent numbers that DO appear in litigation (e.g., 5,999,908 in the Lodsys campaign, 5,999,705 in Burst.com v. Apple). Per your instruction, I am not auto‑correcting or conflating these — they are unrelated patents and unrelated cases.

Case table

Caveats on completeness

• My ability to query paywalled/registration‑gated litigation databases (PACER full dockets, Docket Alarm, RPX Insight, Lex Machina) was limited during this session. The searches that did return results (Unified Patents portal, Google Patents, general web) were negative for this patent number.
• This patent expired on 2016‑10‑02 (anticipated expiration, per the Google Patents legal‑status record). Any enforcement window would have been roughly 1999–2016. I did not find any complaint, ITC Section 337 action, or CAFC appeal in that window.
• Several assignment/security‑interest records exist (SpeechWorks → Nuance; USB AG security agreements; 2016 patent releases), but reassignment records are not litigation and I have not treated them as such.

If you need a definitive negative, I recommend a direct PACER/LEX MACHINA docket query on "5,995,928" and a manual check of the Unified Patents Litigation caselist (https://portal.unifiedpatents.com/litigation/caselist) and RPX Insight, since I could not fully complete those queries here. Based on what I could access, however, the answer is that no known litigation involves US 5,995,928.

I'll verify the canonical "no proceedings" signal against public sources before reporting.

Now let me check for any litigation history or reexamination that could have spawned challenges, and confirm the patent family/assignee chain.

Let me confirm there's no district court assertion or reexamination that could have produced a PTAB record, and check the patent's current status/assignee chain.

PTAB Proceedings Report — U.S. Patent No. 5,995,928

Patent: Method and apparatus for continuous spelling speech recognition with early identification
App. No.: 08/720,554 · Filed: 1996-10-02 · Granted: 1999-11-30
Inventors: John N. Nguyen, Matthew T. Marx · Original assignee: SpeechWorks International, Inc.
Status: Expired — Lifetime (anticipated expiration 2016-10-02)

Proceedings overview

Total AIA trial proceedings on file: 0 (zero). The USPTO Open Data Portal returns no IPR, PGR, or CBM proceeding for US 5,995,928 as of the most recent ingest, and my independent searches surfaced no proceeding the ODP has failed to index. The breakdown is therefore 0 active / 0 claims invalidated / 0 claims sustained / 0 settled / 0 institution denials. The defensive posture this gives a defendant is neither "hardened patent" nor "dead claims" — it is "no live target." The patent expired on 2016-10-02, its full 20-year term having run from the 1996-10-02 filing date, so the practical question is no longer validity but whether any assertable past-infringement window remains (see Strategic summary).

Search-integrity note — do not confuse this patent with its near-namesakes. Searches for "599" patent PTAB activity against a speech/telephony portfolio return several hits that are different patents and must not be attributed to US 5,995,928:

The Davis action is worth flagging only because it is the one piece of litigation touching the '928's original assignee's speech-recognition products. It was a district-court suit on a different patent, produced no PTAB proceeding, and did not involve US 5,995,928 as an asserted patent. Source: Davis v. SpeechWorks Int'l, No. 03-CV-0533S (W.D.N.Y. Sept. 29, 2006).

No proceeding entries follow, because there are none to report. I am not going to manufacture a docket to fill the template; an empty docket is the finding.

Strategic summary

Claim status: all 14 claims UNTESTED at the PTAB — and untestable in any practical sense going forward. Claims 1–14 issued 1999-11-30 and were never canceled, never amended, never subjected to an AIA trial. There is no FWD to quote and no claim-level disposition to report. That is not a clean bill of health — it reflects that the patent was never a litigation-grade asset during the IPR era. SpeechWorks (and its successor in interest, Nuance) never asserted it in a published infringement action that I can locate, and no defensive aggregator — Unified Patents, RPX, AST — ever filed against it. The Unified Patents portal does carry a catalog page for the patent, but a catalog page is not a challenge; there is no Unified-filed petition in the chain.

Assignees and the state of the record. The file shows the assignment chain: Applied Language Technologies (1996-11-21) → SpeechWorks International, Inc. (1999-04-16, by merger) → various security agreements in favor of USB AG / Morgan Stanley, released 2016-05-20 in the ART Advanced Recognition Technologies / Nuance restructuring. Google Patents lists the current assignee of record as SpeechWorks International, Inc.; Unified Patents lists Nuance Communications, Inc. Any demand letter today would need to be checked against the actual USPTO assignment record, because the entity in the chain of title and the entity sending the letter may not match — a § 261 standing problem worth raising immediately if you are contacted.

Estoppel landscape. There is no § 315(e)(2) estoppel, because there was no petitioner and no institution. Nothing is foreclosed. If — contrary to the record — you were facing assertion of this patent, every ground would still be available: § 102 and § 103 on any printed publication or patent predating 1996-10-02, plus § 112 written-description and enablement attacks. The specification is unusually rich in prior-art admissions and in artifacts of the MIT SUMMIT work; the Davis (1991) article and the Marx & Schmandt (1994) paper are cited on the face of the patent, and the SUMMIT/ATIS references (Zue et al.) are admitted state of the art. Those citations are, in effect, a roadmap for a § 102/§ 103 challenge — but you almost certainly will not need it, for the reason below.

The real defense is the calendar, not the art. US 5,995,928 expired 2016-10-02. Under 35 U.S.C. § 286, damages reach back only six years from the filing of a complaint; a complaint filed today (2026-09-28) reaches back to 2020-09-28, which is four years after expiration. There is no period of infringing conduct that is both reachable by the statute of limitations and within the patent term. Section 292 false-marking exposure is separately time-barred for the same reason. In short: no AIA trial activity + expiration = the patent cannot support a live assertion.

Recommended next steps

If you are a defendant and have received a demand letter citing US 5,995,928:

1. Lead with expiration, not invalidity. The patent term ended 2016-10-02. Request the term calculation and the maintenance-fee history; confirm on USPTO Patent Center that the patent is "Expired — Lifetime" (Google Patents so records it). A claim that pre-dates 2020-09-28 is time-barred under § 286; a claim for conduct after 2016-10-02 is barred because the patent had already expired.
2. Check standing and chain of title. Ask the demanding party to produce an unbroken assignment record from Applied Language Technologies through SpeechWorks to the current owner, including the 2016 releases. If the family has been through the ART/Nuance/security-interest restructuring shown on the face of the file, there is a real chance the letter-writer is not the owner of record.
3. Do not spend money on an IPR. Even setting aside expiration, an IPR against an expired patent is largely pointless — the patent owner cannot amend, the claim construction standard is Phillips, and the practical benefit is zero if nothing is assertable.
4. If you nonetheless need an invalidity position on the merits (for example, to defeat a related-patent theory or a counterclaim in a portfolio case), start from the references already cited in the specification — Davis (1991), Marx & Schmandt (1994), Zue et al. (SUMMIT/ATIS) — and the pre-1996 spelling-recognition art. Note that the patent's own § 2 "Prior Art" section admits that touch-tone implicit-disambiguation spelling with early identification was known (Davis), and that the discrete-spoken spelling system with implicit disambiguation and early identification was known (Marx & Schmandt) — admissions that squeeze the novelty of claims 1 and 10 to the "continuous" recitation and the confusability-matrix scoring.

On the absence of PTAB activity as a signal. Normally I would tell you that the absence of IPRs on a well-asserted patent is anomalous and that you should expect one. Here the absence is fully explained and is not a sign that the patent is unusually strong: it expired in 2016, at the tail end of the period in which IPRs became a routine countermeasure, and it does not appear to have been asserted. A patent that was never asserted and is now expired generates no petitions. Treat "no PTAB activity" on this patent as ordinary, not as evidence of hardening.

Confidence and limits. I have high confidence in the zero-proceeding finding as to the ODP ingest and in the 2016-10-02 expiration. I have lower confidence in the completeness of the litigation negative: I found no published district-court assertion of US 5,995,928 specifically, but pre-2000 and sealed dockets are imperfectly indexed, and a defendant-facing investigation should run the patent number through Docket Navigator, Lex Machina, and PACER before relying on it. I found no ex parte or inter partes reexamination either, but my verification search on that point was cut short by a tool limit — confirm it on Patent Center's "Reexam" tab rather than taking my word for it. Also note that the patent's term is close to the outer bound of my certainty only in the sense of fee payment: Google records the patent as "Expired — Lifetime," which is what you would see if all maintenance fees were paid through the 2016 expiration; if any fee was missed, the effective expiration would be earlier, which only helps a defendant.

I'll research the assignment records for US 5995928, focusing on reel/frame numbers and correspondents of record.

Assignment & Ownership Analysis — US 5,995,928

Sourcing note up front. I was able to retrieve the recorded legal-events list for this patent (Google Patents legal-status/assignment events, mirroring the USPTO Assignment Center feed) and to cross-check reel/frame numbers against sibling patents in the same SpeechWorks/Nuance portfolio. I was not able to open the Assignment Center detail pages for US 08/720,554 in this session, so I could not extract every correspondent of record. Where I give a reel/frame, I mark whether it is confirmed for this patent or inferred from a sibling recording in the same portfolio. I do not fabricate reel/frame numbers or attorney names.

Inventors

Unusual patterns: None detectable from the record. Both inventors assigned their rights to the original assignee ~7 weeks after filing (1996-10-02 filing; 1996-11-21 execution) — the normal "file first, paper the assignment second" pattern. I found no evidence that either inventor departed the assignee within 12 months of filing, and no evidence of a fire-sale following inventor exit; I simply have no data either way, so I am not calling it.

Original assignee

Applied Language Technologies, Inc. (Boston, MA), named on the issued patent; subsequently renamed SpeechWorks International, Inc. (merger/change of name recorded 1999-04-16, effective 1998-11-24).

• Line of business: Telephone/embedded speech recognition engines — the "sit-on-a-phone-line" recognizer, sold OEM to carriers and voice portals, not a consumer product.
• Did they ship a product embodying the claims? Yes. The SpeechWorks OpenSpeech Recognizer (OSR) engine line, plus the Speechify text-to-speech product acquired with Eloquent Technology (2000). Deployed on FedEx, Amtrak and Thrifty rental-car caller lines. The claim-1 "spelling engine with confusability matrix" fallback maps to the "spell that name" barge-in behavior these applications used.
• Current status: Acquired. SpeechWorks went public in 2000, was acquired by ScanSoft (announced 2003-04-24, closed 2003-08-11, ~$175.5M in stock), and ScanSoft renamed itself Nuance Communications in 2005. Nuance was acquired by Microsoft for ~$19.7B (announced 2021, closed 2022). The SpeechWorks/Nuance patent estate therefore sits inside Microsoft today. (Sources: Google Patents family/assignment data on this patent; SpeechWorks corporate history at https://rip.so/speechworks.html; SEC filing at https://www.sec.gov/Archives/edgar/vprr/0203/02031416.pdf.)

Assignment timeline

Chronological, as recorded against US 08/720,554 / US 5,995,928. Reel/frame shown where confirmed; where marked "(inferred)," the reel/frame comes from a sibling patent recorded in the same portfolio-wide recording and should be verified in Assignment Center.

• 1996-11-21 (executed 1996-11-21) / recorded 1996-11-21 — Reel not retrieved for this patent

  • Conveyance: Assignment of Assignors' Interest
  • Assignor: Matthew T. Marx; John N. Nguyen
  • Assignee: Applied Language Technologies (Massachusetts)
  • Correspondent: not retrieved
  • Context: Original assignment — inventors assign to their employer; standard file-then-paper.

• 1998-11-24 (effective) / recorded 1999-04-16 — Reel 011502/0869 (inferred from sibling US 6,405,170 B1; the merger recording is portfolio-wide)

  • Conveyance: Merger and Change of Name
  • Assignor: Applied Language Technologies, Inc.
  • Assignee: SpeechWorks International, Inc.
  • Correspondent: not retrieved
  • Context: Internal reorg / change of name only — corporate renaming, not an arm's-length sale.

• 2006-03-31 (effective) / recorded 2006-04-07 — Reel 017435/0199 (confirmed for this patent: the 2016 release expressly cites "REEL:017435/FRAME:0199")

  • Conveyance: Security Agreement (grant of security interest, not a title transfer)
  • Assignor: Nuance Communications, Inc. → note the grantor is already Nuance, evidencing the unrecorded-on-this-page ScanSoft acquisition (2003) and ScanSoft→Nuance rename (2005)
  • Assignee: USB AG, Stamford Branch (Connecticut) — a lender, not an IP buyer
  • Correspondent: not retrieved
  • Context: Securitization — collateral pledge of the patent estate to secure credit.

• 2006-03-31 (effective) / recorded 2006-08-24 — Reel 018160/0909 (confirmed for this patent: the 2016 release expressly cites "REEL:018160/FRAME:0909")

  • Conveyance: Security Agreement (second pledge)
  • Assignor: Nuance Communications, Inc.
  • Assignee: USB AG, Stamford Branch (Connecticut)
  • Correspondent: not retrieved
  • Context: Securitization — layered collateral pledge.

• 2016-05-20 — Reel 038770/0824 (inferred from sibling US 5,179,627 / US 6,360,237; release of 017435/0199)

  • Conveyance: Patent Release
  • Assignor: Morgan Stanley Senior Funding, Inc., as Administrative Agent (successor-in-interest to the USB AG collateral position)
  • Assignee / released party: Nuance Communications, Inc. et al. (grantor group includes SpeechWorks International, Inc., ScanSoft, Inc., Dictaphone, Telelogue, DSP Inc., ART Advanced Recognition Technologies, and others)
  • Correspondent: not retrieved
  • Context: Securitization unwind — lien released; title never left the operating company.

• 2016-05-20 — Reel 038770/0869 (inferred; release of 018160/0909)

  • Conveyance: Patent Release
  • Assignor: Morgan Stanley Senior Funding, Inc., as Administrative Agent
  • Assignee / released party: Same Nuance grantor group, incl. SpeechWorks International, Inc.
  • Correspondent: not retrieved
  • Context: Securitization unwind — second lien released.

Gaps I could not close: (a) the reel/frame for the 1996 inventorship assignment; (b) whether the 2003 SpeechWorks→ScanSoft acquisition and the 2005 ScanSoft→Nuance name change generated separate recorded assignments against this application (the grantor field on the 2006 security agreements implies they did, but the events list for this patent does not surface them); (c) every correspondent of record.

Cross-reference to earlier sections: the prior summary stated the merger was "recorded 1999-04-16" — consistent, and this section adds the effective date 1998-11-24 from the portfolio merger recording. No contradiction.

Timeline diagram

timeline
    title Ownership of US 5995928
    1996 : Filed by Applied Language Technologies
         : Inventors assign to employer
    1998 : ALTech merges into SpeechWorks
    1999 : Merger change of name recorded
    2003 : SpeechWorks acquired by ScanSoft
    2005 : ScanSoft renamed Nuance
    2006 : Patent pledged to USB AG as collateral
         : Second security agreement recorded
    2016 : Morgan Stanley releases both liens
         : Patent term expires 2016-10-02
    2022 : Nuance acquired by Microsoft

NPE / troll-pattern signals

1. Shell-entity transfer — Not present. Every recorded conveyance is to or from an operating company, a successor by merger, or a bank acting as secured lender. No "IP / Holdings / Licensing / Ventures" assignee anywhere in the chain. The 2006 grantee (USB AG, Stamford Branch) is a collateral agent, not an asserting shell.

2. Known asserter in the chain — Not present. No Acacia, Marathon, IV, IPNav, Wi-LAN, Conversant/Mosaid, Vringo, Pendrell, Round Rock, MPHJ, Lumen View, Spangenberg entity, or any Unified Patents / RPX high-frequency plaintiff appears as an assignee. The Unified Patents page for US-5995928-A shows no litigation entries (https://portal.unifiedpatents.com/patents/patent/US-5995928-A).

3. Repeat correspondent across the chain — Unclear. I could not retrieve any correspondent of record for this patent. The pattern I can see — the same reel/frame recordings (011502/0869, 017435/0199, 018160/0909, 038770/0824-0869) applied across the entire SpeechWorks/Nuance portfolio — indicates a single portfolio-wide recording agent, which is normal for a large operating company and is not by itself an NPE tell. Flag for verification, not a finding.

4. Cascading transfers through chained LLCs in <24 months — Not present. The chain spans 1996→2022 and consists of operating-company mergers/acquisitions, not rapid LLC-to-LLC hops.

5. Pre-litigation transfer — Not present. No suit on this patent was found (see the Litigation section), so no assignment can be time-keyed to litigation. The patent expired 2016-10-02.

6. Bankruptcy fire-sale — Not present. SpeechWorks, ScanSoft and Nuance were solvent; the 2003 and 2022 transactions were stock/cash acquisitions, and the 2006 events are a secured financing, not a Chapter 7/11 asset sale.

7. Privateering — Not present. No transfer to an NPE asserting on a sponsoring operating company's behalf. The 2006 security interests were released back to Nuance (reels 038770/0824 and 038770/0869), confirming Nuance retained title rather than farming it out.

8. Defensive aggregator (anti-NPE) — Not present. The chain does not terminate at RPX, AST, LOT, Unified Patents, or OIN.

Verdict

Defensive / non-asserting — with an explicit definitional caveat.

The ownership chain is a clean operating-company succession: inventors → Applied Language Technologies (1996-11-21) → merged into SpeechWorks International (eff. 1998-11-24, recorded 1999-04-16) → into the ScanSoft/Nuance line by 2006 → Microsoft (2022), interrupted only by two security-agreement pledges to USB AG (reels 017435/0199 and 018160/0909, recorded 2006-04-07 and 2006-08-24) that were released back to Nuance by Morgan Stanley on 2016-05-20 (reels 038770/0824 and 038770/0869). Every conveyance traces to an operating company or its lender; there is no NPE, LLC-shell, cascading-transfer, or pre-litigation signal, and the patent expired 2016-10-02. Caveat: the "Defensive / non-asserting" label is defined in the rubric as a chain ending at a defensive aggregator (RPX/AST/LOT/etc.), which does not strictly apply here — the chain ends at a large operating company (Microsoft/Nuance) that simply never asserted this patent. Among the mandated options, "Defensive / non-asserting" is the closest fit because the patent is genuinely non-asserted and the NPE signals are all negative; "Operating-company assertion" does not fit (no suit), and "Insufficient data" does not fit (records exist). If the rubric requires a defensive-aggregator terminus, the correct answer would be "no NPE verdict — operating-company ownership."

Verification links:

• USPTO Assignment Center (search patent number 5995928 / application 08/720,554): https://assignment.uspto.gov/patent/index.html
• USPTO Patent Assignment Search: https://assignmentcenter.uspto.gov/
• Unified Patents portal (no litigation listed): https://portal.unifiedpatents.com/patents/patent/US-5995928-A
• Google Patents legal events for this patent (source of the confirmed 2016 release reel/frame citations): https://patents.google.com/patent/US5995928/en

Open items to close with a primary-source docket check: (1) reel/frame and correspondent for the 1996-11-21 assignment; (2) correspondents of record for reels 011502/0869, 017435/0199, 018160/0909, 038770/0824 and 038770/0869; (3) confirmation of any separately recorded ScanSoft/SpeechWorks acquisition and ScanSoft→Nuance name-change entries for application 08/720,554.

I'll research the prior-art citations for US 5,995,928 from authoritative patent sources.

I have partial front-page reference data. Let me pull the full "References Cited" list and verify each cited patent's details.

Prior Art of US 5,995,928 — Cited References and §102 Mapping

Sourcing note (please read first)

I could not query the USPTO database (PatentCenter/PAIR) or the file wrapper directly in this session — no tool access to those portals. What I could reach and did use:

• The printed front page of the '928 patent (the "References Cited" / U.S. Patent Documents block), via the issued-PDF text at https://patentimages.storage.googleapis.com/cd/b1/44/584708c631b4d3/US5995928.pdf
• Google Patents https://patents.google.com/patent/US5995928/en
• Unified Patents https://portal.unifiedpatents.com/patents/patent/US-5995928-A (its prior-art/reference listing and Non-Patent Literature (18))
• Individual verification searches for the most material references (e.g., US 5,677,990)

Because I could not open the file wrapper, I cannot tell which references were examiner-cited versus applicant-submitted in an IDS, nor what rejections/allowances the examiner made. Where I could not verify a title independently, I say so rather than guessing.

Statutory frame: US 5,995,928 has no priority claim — filing/priority = 1996-10-02. It is therefore governed by pre-AIA 35 U.S.C. § 102. The § 102(b) bar date is 1995-10-02. U.S. patents that issued after 1996-10-02 are available only as § 102(e) art (their U.S. filing dates predate 1996-10-02), and as § 103 support.

Cross-reference flag (contradiction): the Unified Patents portal renders the '928 priority date as 1996-10-01 in several listings, while Google Patents and the assignment record give 1996-10-02 (consistent with your previously-generated summary). I treat 1996-10-02 as authoritative per the front page and prior analysis, and flag the portal's "10-01" as a data-entry discrepancy, not a correction.

Group A — U.S. patent documents printed on the '928 front page (highest confidence)

These are the U.S. Patent Documents actually printed under "(56) References Cited." Dates are as printed (month/year); inventor surname as printed.

A16 expanded — US 5,677,990 (Junqua), the closest cited art

• Full citation: US 5,677,990, Jean-Claude Junqua, "System and Method Using N-Best Strategy for Real Time Recognition of Continuously Spelled Names," issued 1997-10-14; U.S. filing 1995-05-04; assignee Panasonic Technologies, Inc. / Matsushita. (Verified: https://www.freepatentsonline.com/5677990.html; PDF https://patentimages.storage.googleapis.com/be/b4/1c/b69a1e03190042/US5677990.pdf.)
• Disclosure: Recognition of continuously spelled names uttered as a letter sequence (expressly for telephone directory lookup). A first recognizer applies an n-gram letter grammar (bigram letter pairs) and produces N-best letter-sequence hypotheses with probability scores and adaptive path pruning; a dynamic time warping (DTW) alignment compares the N-best hypotheses against a name dictionary to select N-best name candidates; a dynamic grammar then constrains a second recognizer to pick the best name. The specification states the goal is real-time operation over a telephone channel and addresses letter confusability (E-set etc.).
• §102 assessment:
  • Claims 10 and 14 (method/computer-implemented method: receive continuously uttered letters → analyze → periodically produce hypothesized letter strings → compare against a preselected vocabulary until a best match): this reference is a serious §102(e) candidate, and on its face discloses every step except it may not explicitly recite periodic partial-string output before completion in the same terms.
  • Claims 1 and 2 (spelling engine with a confusability matrix + "one word that best matches"): Junqua's Table I reports a "Confusability" metric, but that is a name-dictionary confusability figure, not a letter-pair confusability matrix as claim 1 requires. So Junqua is better characterized as §103 art against claim 1 (combined with, e.g., the Marx/Schmandt confusability list — see C2), rather than a clean §102(e) anticipant.
  • Flag: because Junqua issued 1997-10-14, it can only be §102(e) art; it cannot be §102(a)/(b).

Group B — Additional U.S. references in the Unified Patents reference listing (medium confidence as cited art)

These appear in the prior-art/reference list on the Unified Patents page for US-5995928-A but were not visible in the PDF front-page excerpt I retrieved. Treat the "cited" status as medium confidence; the titles/dates are from that listing.

Not to be confused: US 5,748,840 (La Rue, "Methods and apparatus for improving the reliability of recognizing words in a large database when the words are spelled or spoken," filed 1995-05-09) surfaced in my searches as related technology in the same field, but I did not confirm it as a cited reference of the '928 patent, so I am not listing it as '928 prior art. It would nonetheless be a §102(e)-era reference worth checking in the file wrapper.

Group C — Non-Patent Literature (18 items; the most legally important group)

(The remaining NPL slots on Unified's list of 18 are additional SUMMIT/recognizer papers of the same character as C3–C7; I could not enumerate all 18 titles from the sources I reached, and I will not invent them.)

Where the §102 risk actually concentrates

1. Independent claims 1 / 10 / 14 depend on the combination of (a) continuous spelling, (b) a real-time/partial recognizer hypothesis stream, and (c) a dictionary/disambiguation comparison. No single cited U.S. patent I found cleanly discloses all three as continuous spelling. US 5,677,990 (A16, Junqua) comes closest — it is the single most dangerous § 102(e) reference for claims 10 and 14, and a strong § 103 anchor against claim 1. US 5,125,022 (A7) and US 5,638,425 (A15) are the next most relevant for the spoken-letter/vocabulary-matching element.
2. Claim 1's "confusability matrix" is the softest element. The applicant's own Background admits the Marx & Schmandt confusability list (C2) as prior art, and Davis (C1) as the source of vocabulary disambiguation + early identification. Combined with Junqua's continuously-spelled-names pipeline, the confusability-matrix limitation is exposed to a § 103 attack even though no cited reference anticipates it alone.
3. Claims 6 and 8–9 (early identification) are squarely pre-empted in concept by C1 and C2, both of which the specification itself describes as providing early identification. Those are § 102(b) admissions of record.
4. Claims 3–4 and 11 (node grid + DP scoring + insertion/deletion/confusability transition costs) rest on textbook DP/DTW (C9) plus the applicant's own confusability approach — again a § 103-vulnerable combination rather than a single-reference anticipation.
5. Claims 5 and 12 (threshold + margin "delta" criteria) were not found anticipated by any cited reference I reviewed; this appears to be the most defensible subject matter on the face of the cited art.

Explicit limitations on this analysis

• I did not access USPTO PatentCenter/PAIR, so I cannot separate examiner-cited from applicant-cited (IDS) art, cannot cite the examiner's actual § 102 rejections or the reasons for allowance, and cannot confirm whether any of Group B was formally listed under "(56)."
• Titles/dates for A1–A6, A10–A12 and B4–B5 are not independently verified in this session; only the numbers/dates/inventor surnames come from the printed front page, and I have flagged the unverified ones rather than supplying titles from memory.
• § 102 characterizations above are technical relevance mappings, not legal conclusions; anticipation additionally requires that a single reference disclose every element as arranged in the claim, which I could only assess in detail for US 5,677,990.
• One data conflict is flagged above (1996-10-01 vs. 1996-10-02 priority date on the Unified Patents portal).

I'll ground this obviousness analysis in the actual prior-art references of record for US 5,995,928. Let me pull the "References Cited" section and verify the key non-patent references.

Let me pull the examiner-cited U.S. patent references and verify the SUMMIT/segment-based art.

Let me verify the examiner-cited U.S. references' subject matter so I can map them accurately.

Obviousness Analysis — U.S. Patent No. 5,995,928 under 35 U.S.C. § 103

Building on the previously generated bibliographic summary and litigation summary (which I do not repeat here). No contradictions with those sections were found. One cross-reference point matters for this analysis: the litigation section found no district-court, ITC, or Federal Circuit record for this patent, which means there is no adjudicated validity record, no IPR estoppel, and no litigation-tested objective-indicia record to draw on.

1. Sourcing caveats (read before the analysis)

Per the operating rules, I am flagging confidence levels explicitly:

This means the primary § 103 grounds below rest on references whose content I can actually attribute (Marx & Schmandt 1994; Davis 1991; Zue/SUMMIT; Rabiner & Juang 1993), plus Meador III. The remaining examiner-cited patents are treated as secondary/alternative art only, with an explicit limitation.

2. Legal framework and the person of ordinary skill

Pre-AIA § 103 governs (filed 1996-10-02), as construed under Graham v. John Deere Co., 383 U.S. 1 (1966) and KSR Int'l Co. v. Teleflex Inc., 550 U.S. 398 (2007). KSR applies to all cases on review, and it controls a key point here: no express teaching, suggestion, or motivation in the references is required. It is enough that the combination was "obvious to try" where there was "a finite number of identified, predictable solutions," or that the combination is "of familiar elements according to known methods ... yield[ing] predictable results." Id. at 421. See also MPEP § 2143 rationales (A)–(G), including (C) "use of known technique to improve similar devices in the same way" and (F) "obvious to try."

POSITA: a person with an M.S. (or equivalent experience) in EE/CS plus 2+ years in speech recognition, or a B.S. with 4–5 years, familiar with (i) Viterbi/DP search over segment lattices, (ii) dynamic time warping and edit-distance string alignment, and (iii) telephony IVR/ASR applications. The specification's own citations (Rabiner & Juang; the SUMMIT papers; Davis) define this art.

Critical admission in the specification. The '928 patent repeatedly characterizes its own building blocks as conventional. The patent states its recognition engine "may be a segment-based recognition engine such as the SUMMIT system or any other commercial or research speech recognition engine which can output a hypothesis of partial spellings as a user speaks" (col. describing SUMMIT), and that "conventional speech decoders typically wait for the end of speech ... then conduct a backward pass" while engines "typically perform the matching process in forward mode until the entire utterance has been received." A patentee's specification admissions about what the art already does are prior-art-equivalent for § 103 purposes. This substantially narrows the disputed ground to: the spelling engine's scoring machinery, and the timing used to trigger it.

3. The prior-art record available on this page

(a) Non-patent literature (verified content):

(b) U.S. patent documents of record (front page): 4,866,778 (Baker); 4,947,438 (Paeseler); 5,005,203 (Ney); 5,018,201 (Sugawara); 5,027,406 (Roberts et al.); 5,105,465 (Koyama); 5,125,022 (Hunt et al.); 5,131,043 (Fujii et al.); 5,202,952 (Gillick et al.); 5,204,894 (Darden); 5,233,681 (Bahl et al.); 5,454,063 (Rossides); 5,526,463 (Gillick et al.); 5,638,425 (Meador III et al.).

Of these, I verified only Meador III: "Automated directory assistance system using word recognition and phoneme processing method" (Bell Atlantic Network Services; filed 1992-12-17; issued 1997-06-10), disclosing "multiple speech recognition devices comprising a word recognition device, a phoneme recognition device, and an alphabet recognition device" plus a voice processing unit in a telephone directory-assistance context. I could not verify the disclosures of the other thirteen and will not attribute teachings to them.

4. Element mapping — the independent claims

Claim 1 (system)

Claim 10 / Claim 14 (method; claim 14 adds the output step)

All steps mirror claim 1, with "process[ing] the letters into a speech signal ... compatible with a speech recognition engine" (SUMMIT front end per the spec: A/D, framing, MFCC — and the patent itself calls these steps conventional) and "until a best match is obtained." Claim 14's terminal "outputting the vocabulary word, before the word has been completely spoken" is verbatim the early-identification function of Davis and Marx & Schmandt.

5. The three strongest § 103 grounds

Ground 1 (primary): Marx & Schmandt 1994 + Davis 1991 + Rabiner & Juang 1993

The combination. Marx & Schmandt teach the complete functional architecture of the back half of claim 1 — spoken letter recognition → confusability mapping of hypothesized letters to actually-spoken letters → comparison of candidate letter sequences against an allowed-word list → narrowing to a single word and identifying it before spelling is complete. Davis supplies the vocabulary-list/multi-candidate bookkeeping and the express articulation of early identification as a performance feature. Rabiner & Juang supply the one genuinely missing piece: a DP string-alignment formulation in which substitutions are scored by a confusion weight and insertions/deletions are scored by transition penalties — i.e., the "node grid" of claims 3–4 with diagonal/horizontal/vertical moves.

Why the combination is obvious (KSR / MPEP 2143):

1. Same problem, same field (rationale: motivation from the nature of the problem). All three address reliable recognition of spelled input as an ASR fallback. Pro-Mold & Tool Co. v. Great Lakes Plastics, 75 F.3d 1568, 1573 (Fed. Cir. 1996). The '928 spec itself frames the invention as fixing the discrete system's drawback.
2. The '928 patent admits the discrete system's limitation was a speed/usability problem, not a feasibility problem — "because the user must wait for the system to prompt him or her for each letter, the system appears slow and time-consuming"; and "a system that seems time-consuming or slow may be avoided by a user." Designing for continuous input to remove per-letter prompting latency is the predictable use of a known recognition technique to improve a known device ready for improvement (MPEP 2143(C)). KSR: the improvement is "of familiar elements according to known methods" yielding "predictable results."
3. The insertion/deletion problem was the expected, known obstacle with a known solution. The '928 background concedes continuous spelling causes "the recognizer [to] erroneously insert and delete letters in its hypothesis." The standard, well-known cure for exactly that — DTW/DP alignment with insertion/deletion/substitution costs — is what Rabiner & Juang, expressly cited by the patent, set out at pp. 204–241. There is no discovery here; there is an application of a textbook technique to an identified problem. That is a textbook "obvious to try" case with a finite number of identified, predictable solutions (KSR, 550 U.S. at 421).
4. The confusability-matrix element is disclosed or, at most, an obvious formalization. Claim 1 recites "a confusability matrix representing the confusability between each hypothesized letter and each letter of each word within the vocabulary list." Marx & Schmandt's per-letter misrecognition list (v → {b, d, e, p, v, z}) is a confusion table over the same letter universe that makes up the words in the vocabulary. Organizing it as a two-dimensional letter-pair matrix is a data-structure choice a POSITA would make as a matter of course when the alignment moves diagonally through a grid — indeed, the '928 spec's own FIG. 3 matrix is populated with the kind of values (A→A = +10; A→S = −10) that any empirical confusion study yields. Cf. In re Boesch / In re Antonie (structural/data-format changes producing no unexpected result are within the skill of the art).
5. The timing element is a design choice with an admitted trade-off. Emitting "an updated string ... before the word has been completely spoken" is achieved simply by reading out the forward-mode decoder's currently-best path, which the spec concedes engines already compute ("the decoder ... is run in the forward direction to find the best match"). The spec also concedes that performing a full forward–backward run "for every partial hypothesis would ... consume so much system resources ... as to be impracticable" — i.e., the choice of forward-only, mid-utterance output was a recognized, predictable engineering trade-off, not an inventive contribution.

Prosecution-history/art-recognition corroboration. A later PCT search report (WO 2017084334 A1) cites US 5,995,928 A itself as "Y" (obviousness) art against claims 1–18 of a subsequent application, and the EP3394852 supplementary search report cites a 2014 thesis as "X" art for spelling-recognition claims — evidence that the claimed subject matter is treated in the art as the conventional way to build continuous-spelling recognition. (Both documents post-date the '928 filing, so they are not prior art; they are cited only as evidence of what the field regards as routine.)

Ground 2: Meador III (US 5,638,425) + Davis 1991 + Rabiner & Juang 1993

Meador III, verified above, discloses an automated telephone directory-assistance system with a word recognition device, a phoneme recognition device, and an alphabet recognition device. Two things follow. First, the commercial context of the '928 patent — telephony fallback for names — was squarely known. Second, an "alphabet recognition device" in a directory-assistance system is recognition of letters for spelled name input, i.e., the front half of claims 1/10/14. Adding Davis's early-identification/vocabulary-disambiguation logic (which Meador's directory-assistance system needs for "which subscriber name?") and Rabiner & Juang's DP alignment yields every claim-1 element. Motivation: Meador and Davis are directed to the identical problem — resolving a user-supplied spelled word (or keystroke string) against a constrained name/word list, where latency and accuracy drive the design. Pro-Mold.

Ground 3: Zue/SUMMIT + Marx & Schmandt + Davis + Rabiner & Juang

For claims 10 and 14 (and the recognizer half of claim 1), the SUMMIT papers supply a complete, admitted available engine that "can output a hypothesis of partial spellings as a user speaks," including segment lattice construction, phoneme classification, forward Viterbi decoding against a letter lexicon, and a bigram language model over letters. The '928 spec concedes that any such commercial or research engine may be used. The combination adds the Marx & Schmandt/Davis word-list disambiguation + early identification and the Rabiner & Juang DP string match. Reason to combine: the '928 spec itself states the commercial goal (early identification "often before the user has finished spelling"), and the front-end processing steps are conceded to be conventional digital signal processing.

6. Dependent claims 2–9 and 11–13

7. The strongest counter-arguments a patent owner would raise (and why they likely fail)

1. "The art taught away from continuous spelling." Marx & Schmandt chose discrete prompt-per-letter precisely because it "avoids any confusion over how many letters were spoken," and the '928 background stresses that continuous spelling is "a difficult problem" (boundary ambiguity, insertions, deletions). This is the best non-obviousness argument.
   Rebuttal: Identifying a difficulty is not teaching away. Teaching away requires that the art "criticize, discredit, or otherwise discourage" the claimed solution such that a POSITA would not have pursued it. In re Fulton, 391 F.3d 1195, 1201 (Fed. Cir. 2004); DePuy Spine, Inc. v. Medtronic Sofamor Danek, Inc., 567 F.3d 1314, 1327 (Fed. Cir. 2009). Marx & Schmandt merely note a design trade-off — they do not assert that continuous spelling cannot work, and Davis had already achieved continuous-input early identification in the telephone-keypad domain. Moreover, the specific difficulty identified (insertion/deletion) is precisely what DP alignment was known to solve, so the art pointed toward the claimed solution.
2. "Hindsight: the references come from different domains (keystroke disambiguation, discrete spelling, DTW)." In re Rouffet / KSR hindsight concerns.
   Rebuttal: The references are all in the same field of endeavor (speech recognition and spelled-input disambiguation), all published in the same small community (AVIOS proceedings, ICASSP), and Davis and Marx & Schmandt share the identical objective — early identification from a candidate-letter-sequence search against a word list. Combining a known DP string-matcher with a known candidate-search architecture is not hindsight; it is routine engineering.
3. "The examiner allowed the claims over this exact art." True — the 14 U.S. references and the NPL were of record, and the patent issued with 14 claims in Nov. 1999. This is a real, non-trivial obstacle to a § 103 challenge (a reexamination/IPR petitioner would have to show the examiner erred, e.g., by not appreciating the Marx & Schmandt confusability-list teaching as the claim-1 "confusability matrix," or by not articulating a KSR motivation). It is not, however, a legal bar, and examiner allowance carries no presumption of validity beyond the normal § 282 presumption.
4. Claim-construction escape hatch. A patent owner could argue "confusability matrix" requires a 2-D letter-pair matrix (per FIG. 3) and "each letter of each word within the vocabulary list" requires conditioning on vocabulary membership, so Marx & Schmandt's 1-D per-letter list does not disclose it.
   Assessment: This is the most defensible position for the patent owner, and it should be litigated on the intrinsic record. But even if the construction holds, the difference is a data-representation choice with no unexpected result — classic In re Antonie optimization — so it supports, at most, non-anticipation, not non-obviousness.

8. Secondary considerations

• No objective indicia are of record here. There is no litigation in which validity was tested (per the litigation section of this analysis), no IPR, and no evidence of unexpected results, long-felt need, failure of others, or copying available in this record.
• Commercial success is plausible in principle: the assignee chain (Applied Language Technologies → SpeechWorks → ScanSoft/Nuance) sold telephony ASR products in the relevant window. But under Polaris Industries, Inc. v. Arctic Cat, Inc., 882 F.3d 1056 (Fed. Cir. 2018), the patentee would have to show the success is coextensive with the claimed features; the claimed spelling engine was plainly one component of a broader product. No such nexus evidence exists in this record.
• Unexpected results would need to be tied to a specific claim limitation (e.g., the threshold/delta criteria). The spec's assertion that values "have been shown through experimentation ... to achieve a high recognition accuracy" is the opposite of an unexpected-result showing — it is routine optimization.

9. Bottom line

On the record available on this page, claims 1, 10, and 14 are prima facie obvious over Marx & Schmandt 1994 in view of Davis 1991 and Rabiner & Juang 1993 (with Zue/SUMMIT as the admitted recognition front end), and alternatively over US 5,638,425 (Meador III) in view of Davis and Rabiner & Juang. Dependent claims 2–9 and 11–13 add only (i) the standard DTW/Viterbi recurrence the art already used, (ii) threshold-plus-margin confidence criteria and selected numeric values constituting optimization of result-effective variables, and (iii) an IVR output modality conceded to be conventional. The strongest single hurdle is not the technical substance but the procedural posture: these references were of record and the examiner allowed the claims anyway, so a challenger must show the examiner failed to appreciate the Marx & Schmandt confusability-list teaching or failed to articulate a KSR motivation.

Two explicit limitations on this opinion: (1) I did not verify the subject matter of 13 of the 14 examiner-cited U.S. patents, so any ground built on those references (e.g., 4,866,778 Baker; 5,233,681 Bahl et al.; 5,202,952 and 5,526,463 Gillick et al.) is not established here and would require inspecting those documents; and (2) I could not retrieve the '928 file wrapper / examiner's reasons for allowance, so I cannot say whether the examiner expressly considered Marx & Schmandt's confusability list against claim 1's "confusability matrix" limitation.

Key sources: https://patents.google.com/patent/US5995928/en · https://portal.unifiedpatents.com/patents/patent/US-5995928-A · https://patentimages.storage.googleapis.com/cd/b1/44/584708c631b4d3/US5995928.pdf · https://patentimages.storage.googleapis.com/93/8a/38/0da5802d1f985b/US5638425.pdf · https://patentimages.storage.googleapis.com/e5/55/40/10dc454b753ee7/WO2017084334A1.pdf · https://insight.rpxcorp.com/patent/[US5995928A](/patent/US5995928A)