US 10909987

Patent Overview: US10909987B2 — "Hotword detection on multiple devices"

Title: Hotword detection on multiple devices
Patent number: US10909987B2
Application number: US16/553,883
Inventor: Matthew Sharifi
Current assignee: Google LLC (original assignee also listed as Google LLC)
Priority date: 2014-10-09
Filing date: 2019-08-28
Publication date: 2021-02-02 (as US10909987B2; earlier publication US20200058306A1 on 2020-02-20)
Legal status: Active
Anticipated expiration: 2035-04-01
Source: https://patents.google.com/patent/[US10909987](/patent/US10909987)/en

Purpose: The specification relates generally to speech recognition. In a speech-enabled environment (e.g., home, workplace, school) implemented via a network of connected microphone devices, the system must discern when an utterance is directed at it rather than at a person present. One approach uses a hotword—a predetermined word, e.g., "OK computer"—spoken to invoke the system's attention. Utterances take the general form [HOTWORD] [QUERY].

Plain-language scope: A user device receives a spoken utterance, determines whether it includes a hotword, and computes a hotword confidence score indicating the likelihood the hotword was present. The device transmits this score to nearby user devices, which likely received the same utterance; those devices compute and return their own scores. The first device compares the scores: if it has the highest score, it remains active and prepares to process additional audio; if not, it does not process the additional audio.

Described method actions include: receiving audio data corresponding to an utterance; determining a first value for hotword likelihood; receiving a second value determined by a second computing device; comparing the values; and, based on the comparison, initiating speech recognition processing. Further actions include determining the first value satisfies a hotword score threshold; transmitting the first value to the second computing device; and determining an activation state (active or inactive). Transmission/receipt may occur to/from a server, through a local network, or through a short-range radio, and may include device identifiers.

What the excerpt does not establish: It does not provide the full claim set, complete written description, or prior-art references, and the text notes omitted content.

I'll research litigation involving US patent 10909987 across patent litigation databases.

Summary of Findings: Litigation Involving U.S. Patent No. 10,909,987

Bottom line: I could not confirm any litigation (district court or PTAB) in which U.S. Patent No. 10,909,987 ("Hotword detection on multiple devices," Google LLC) is itself an asserted or challenged patent. Below is what the searches did and did not turn up.

What I searched

• Unified Patents Portal — patent page for US-10909987-B2: https://portal.unifiedpatents.com/patents/patent/US-10909987-B2. The page returns bibliographic/portfolio data (Assignee: Google LLC; Inventor: Matthew Sharifi; Application No. 16/553,883; Priority Date 2014-10-08; Grant Date 2021-02-01; Expiration 2035-03-31). No litigation case list is shown on the retrieved record.
• Google Patents — the exemplary page for US10909987B2 (https://patents.google.com/patent/US10909987/en).
• Docket Alarm / CourtListener / general web — searches for the literal number "10,909,987" and "10909987" in complaint and docket contexts.

One important caveat / lead (not a confirmed case)

The Google Patents record for US10909987B2 carries a "Family has litigation" flag with a Darts-IP link (https://patents.darts-ip.com/?family=54330874). This flag is family-level, not patent-level — it indicates that some member of the 48-member patent family (which includes continuations such as US 11,557,299, US 11,915,706, US 12,254,884, and publication US 2025/0191590) has been involved in litigation somewhere. It does not establish that the '987 patent itself was asserted, nor does it identify a plaintiff, defendant, court, or case number. I was unable to resolve the underlying case from the available search results, so I am not reporting it as a case against the '987 patent.

Note on search noise

My queries surfaced several cases containing similar but different patent numbers (e.g., 11,096,039 in Mullen Industries LLC v. Apple Inc., N.D. Cal. No. 3:23-cv-00437; 10,909,xxx drug-trial entries; "10X Genomics" matters). Per your instruction, I have not conflated any of these with U.S. 10,909,987 — none of them involve this patent.

Requested table

Recommended next steps to verify definitively

1. Unified Patents litigation caselist filtered by the patent: https://portal.unifiedpatents.com/litigation/caselist?patents=10909987 (the portal supports a ?patents= filter).
2. USPTO Patent Trial & Appeal Board — search IPR/PGR petitions by patent number at the PTAB E2E / PTAB Data portal.
3. PACER / CourtListener RECAP — full-text docket search for "10,909,987" (Google Patents "Litigation" tab, when populated, mirrors these).
4. Darts-IP family 54330874 — to identify which family member(s) have been litigated, which may or may not include the '987 patent.

Caveat on confidence: The absence of results is a negative finding from the sources I was able to reach; it is possible a case exists that is not indexed in those sources or that data beyond my retrieval step was truncated. I am not asserting that no litigation exists — only that I could not find or confirm any asserting this specific patent number.

I'll start by verifying the structured "no PTAB activity" signal against web sources, then check for family-member proceedings and litigation.

The family IPRs (Sonos v. Google on the '398 and '330 patents) surfaced, but those are sibling patents, not the '987. Let me verify whether the '987 itself is the subject of any AIA trial.

Let me run a final confirmation check on PTAB docket aggregators and the sibling-patent litigation to make sure no '987-specific trial exists.

PTAB / AIA Trial Landscape — U.S. Patent No. 10,909,987 B2 ("Hotword detection on multiple devices," Google LLC)

Proceedings overview

Total AIA trial proceedings on file: 0. The USPTO Open Data Portal returns no IPR, PGR, or CBM proceeding naming U.S. Patent No. 10,909,987 B2, and my independent web checks (PTAB decision portals, PTAB E2E-adjacent docket aggregators, CourtListener, and the PTAB/Fed. Cir. coverage of the Sonos–Google hotword dispute) surfaced no petition, institution decision, FWD, or appeal directed at this patent. The breakdown by status is therefore nil across every category — active (0), claims invalidated (0), claims sustained (0), settled (0), institution denied (0). The correct defensive gloss is not "the patent survived IPRs and is hardened" and not "the claims are canceled." It is: all 21 claims of the '987 are UNTESTED at the PTAB. A defendant faces a patent whose validity has never been adjudicated in an AIA trial — which cuts both ways (no estoppel shield for you, but also no precedent cancelling the claims you are accused of infringing).

---

Proceedings on U.S. Patent No. 10,909,987 B2

None. No AIA trial proceeding exists to report. I am not going to populate the per-proceeding template with fabricated case numbers.

---

Adjacent family activity (expressly NOT proceedings on the '987)

Because the '987 is a member of the same § 119/§ 120 family as two Google patents that were litigated at the PTAB, the following is highly relevant background for anyone evaluating the '987 — but every item below concerns a different patent with different claims. Treat it as intelligence, not as estoppel or as a claim-level holding on the '987.

Sonos, Inc. v. Google LLC — IPRs on U.S. Pat. Nos. 10,134,398 and 10,593,330

• Type: Inter Partes Review (two proceedings, one per patent)
• Filed: Petitions reported as filed 2022-09-29 and 2022-10-27 (per the 4IP Council Sonos–Google litigation survey, Table 4.7); docket aggregators surface the pair as IPR2023-00118 and IPR2023-00119 — I cannot independently verify those case numbers from a primary PTAB source, so confirm on PTAB E2E before citing them.
• Status: Final Written Decisions issued 2024-05-15; both FWDs reversed and remanded by the Federal Circuit on 2026-06-09.
• Claims at issue (as recited in the CAFC opinion): claims 1–5, 7–13, and 15–20 of the '398 patent and claims 1–7, 9–15, and 17–18 of the '330 patent.
• Petition grounds: § 102 anticipation and § 103 obviousness over U.S. Pat. No. 8,340,795 (Rosenberger), and Rosenberger in view of U.S. Pub. No. 2014/0163978 (Basye).
• Board holding: claims 1–3, 7–11, and 15–18 of the '398 and claims 1–3, 5, 7, 9–11, 13, 15, and 17–18 of the '330 unpatentable as anticipated by Rosenberger; claims 4, 5, 12–13, and 19–20 of the '398 and claims 4, 6, 12, and 14 of the '330 unpatentable as obvious over Rosenberger + Basye.
• Appeal: Google LLC v. Sonos, Inc., No. 2024-2119 (Fed. Cir. June 9, 2026), nonprecedential, opinion by Chief Judge Moore. The court held the Board's finding that Rosenberger discloses exchanging messages while the device remains in the low power mode was unsupported by substantial evidence, reversed, and remanded. The court also declined to affirm on Sonos's alternative "reduced power consumption" theory because the Board made no factual findings on it.
• Citations: https://law.justia.com/cases/federal/appellate-courts/cafc/24-2119/24-2119-2026-06-09.html · https://www.courtlistener.com/opinion/[10872385](/patent/10872385)/google-llc-v-sonos-inc/
• Defensive value for the '987: Mixed. The Board's 2024 FWDs show that the family's core "confidence-score exchange / low-power coordination" concept can be attacked with pre-2014 art, and Rosenberger remains on the table. But the June 2026 reversal teaches that the Board applied Rosenberger too loosely — a defendant copying the Sonos playbook must land explicit, column-and-line evidence that the reference performs coordination while still in the low-power state, not merely that the reference contains both features somewhere.

ITC Inv. No. 337-TA-1329, Certain Audio Players and Components Thereof (I)

• Google's complaint asserted U.S. Pat. Nos. 7,705,565, 10,593,330, and 10,134,398 — not the '987. Instituted 2022-09-15; the '565 patent was terminated 2022-11-02. The Commission's 2025-05-27 notice granted Google's motion for reconsideration, vacated the 2025-04-08 determination that terminated the investigation with a finding of no violation, and reopened the investigation (target date July 7, 2025) on the ALJ's summary determination of invalidity of the asserted claims.
• Source: https://www.usitc.gov/secretary/fed_reg_notices/337/337_1329_notice05272025_reconsideration.pdf
• Defensive value for the '987: none directly — different patents — but it confirms the '987's siblings were the asserted ones, and it shows how volatile the validity posture of this family has been across forums (PTAB invalid → CAFC revives; ITC invalidity determination → vacated and reopened).

---

Strategic summary

Claim status on the '987: 100% UNTESTED. All 21 claims (independent claims 1, 9, and 16, plus dependents) stand exactly as issued on 2021-02-02, with no PTAB cancellation, no certificate of correction narrowing them, and no AIA-trial estoppel attached to any of them. If a demand letter or complaint against you cites claims 1–21, none of those claims has been invalidated by the Board — and, just as importantly, none has been upheld by the Board either. There is no FWD to quote and no cancellation to point to. (Note: the '987's claim set as issued recites the confidence-score comparison architecture — "determining a first value corresponding to a likelihood that the utterance includes a hotword… receiving a second value… comparing the first value and the second value… initiating speech recognition processing on the audio data" — which is a different claim scope from the "low power mode / on-device hotword detector" limitations that drove the '398/'330 IPRs. Art that killed or nearly killed the '398 and '330 will not automatically map onto the '987's claims; claim-chart the '987 independently.)

Estoppel landscape. Because no IPR or PGR was ever instituted on this patent, § 315(e)(2) estoppel does not bar any ground against the '987 — not for Sonos, not for Google, not for any third party, and not for privies of any prior petitioner. Nor does § 325(e)(2) apply. The Sonos petitions against the sibling '398 and '330 do not create estoppel as to the '987: estoppel under § 315(e)(2) runs to "the patent" and the claims subjected to the instituted IPR, and the '987 is a separate patent with separate claims. Practically, that means a defendant today retains the full universe of § 102/§ 103 prior-art grounds against the '987, including Rosenberger, Basye, and any art that Sonos did or could have raised. The one real constraint is budgetary and discretionary, not statutory: if you file an IPR on the '987, you will be litigating the family's art a second time, and the Board may weigh the parallel district court posture under § 314(a)/Fintiv.

Pattern signals. Three observations. (1) The same petitioner did not file multiple IPRs on this patent — there are zero. (2) The patent owner (Google) litigates its PTAB appeals aggressively: it appealed both FWDs in the sibling IPRs and won a full reversal and remand at the Federal Circuit in June 2026, meaning Google has demonstrated both the will and the appellate track record to defend this family's validity. (3) No defensive aggregator appears in the chain. Unified Patents maintains a public page for US-10909987-B2 (https://portal.unifiedpatents.com/patents/patent/US-10909987-B2) but there is no Unified-filed IPR or PGR on it — the page is a monitoring/rating entry, not a challenge. The only "litigation family" hit is the Darts-IP family flag on the Google Patents page, which reflects the broader Sonos–Google war, not a proceeding against the '987.

---

Recommended next steps

• If you are a defendant facing assertion of the '987: Do not assume any claim is dead. There is no FWD to link to and no canceled claim to quote — the Board has never ruled on this patent. Your invalidity case must be built from scratch against the '987's own claim language, and the strongest lead is the Rosenberger/Basye line that Sonos used against the sibling '398/'330, re-charted against the '987's "first value / second value / comparison" limitations.
• Mind Artis / § 315(b)-adjacent and IPR-vs-district sequencing. With no prior IPR, you are free to file, but expect the patent owner to argue § 314(a)/Fintiv against institution if there is a parallel district court case, and expect it to cite its successful June 2026 appellate defense of the family.
• If you represent Google or another patent owner: the absence of PTAB activity is a genuine asset — there is no adverse FWD, no cancellation certificate, and no petitioner confining itself to an estoppel box. The absence also suggests the '987 has not been a primary target of any competitor's validity campaign, unlike its '330 and '398 siblings.
• If you are looking for confirmation before relying on any of this: pull the patent's "PTAB proceedings" tab on PTAB E2E / the USPTO Open Data Portal and re-check CourtListener for any post-2024 appeal naming the '987. My searches through 2026-10-01 found none. The absence of PTAB activity is itself the signal here: this patent has been in force since 2021 and its siblings have drawn heavy IPR fire, yet no petitioner has ever filed against the '987.

Confidence and caveats. High confidence on the zero-count result (consistent across the structured ODP block and my web checks). Lower confidence on the two sibling IPR case numbers (IPR2023-00118 / -00119) — those come from a third-party docket aggregator and were reported inconsistently alongside a "2022" prefix in the same exhibit; verify on PTAB E2E before citing. The Federal Circuit case number (24-2119), decision date (2026-06-09), and claim-level dispositions of the '398/'330 FWDs are taken verbatim from the opinion text and are reliable. I found no public record of the '987 being asserted in ITC 337-TA-1329 or in the Sonos–Google district court actions; the asserted hotword patents there were the '330 and '398, not the '987.

I'll research the assignment record for US 10,909,987, starting with the authoritative legal events already captured and then verifying against USPTO Assignment Center and related sources.

Let me search for the specific recorded assignment details (reel/frame and correspondent) for this patent and Google's entity-conversion practice.

Assignment & Ownership Reconstruction — US 10,909,987 B2 ("Hotword detection on multiple devices")

Scope note carried from prior sections: the grant is US 10,909,987 B2, App. No. 16/553,883, inventor Matthew Sharifi, original assignee Google Inc., current assignee Google LLC. The chain is a pure continuation family (provisional 62/061,830 → 14/675,932 → 15/088,477 → 15/346,914 → 16/171,495 → 16/553,883 → later 17/137,157). I treat the patent's own CROSS-REFERENCE section and the Google Patents "Legal Events" tab as authoritative, because USPTO Assignment Center's post-2024 interface did not return a machine-readable record for this number through my search path.

---

Inventors

Basis: US 10,909,987 lists a single inventor, Matthew Sharifi. The original provisional (62/061,830, filed 2014-10-09) and the first non-provisional (14/675,932, filed 2015-04-01) were filed while Sharifi was a Google engineer; the application was filed as a Google Inc. application and the recorded inventor-assignment names Google Inc. as assignee (see timeline). No separate corporate assignee is listed for him on the face of the patent.

Unusual-pattern check: Not present. There is no evidence of the inventor departing the assignee, and no mass inventor-assignment event preceding a sale. Sharifi is a prolific, continuing Google inventor — my searches surfaced him as named inventor on later Google hotword/Assistant grants including US 10,262,659, US 10,446,153, US 11,308,934, and US 12,394,416, all currently assigned to Google LLC. A single long-tenured inventor on a still-Google-owned family is the opposite of a fire-sale signature.

---

Original assignee

• Entity on the issued patent: Google LLC, 1600 Amphitheatre Parkway, Mountain View, CA 94043 (originally Google Inc., converted 2017-09-30).
• Product embodying the claims: Yes. The claims cover multi-device hotword arbitration — nearby devices each score a spoken wake word, exchange scores, and only the best scorer activates full ASR. This is the mechanism behind Google's "Hey Google"/"OK Google" wake-word handling across Android phones, Nest/Google Home speakers and displays, Wear OS, and other Assistant-enabled surfaces. Google ships these at scale.
• Primary line of business: search, advertising, cloud, consumer hardware, and the Google Assistant / Android ecosystem.
• Current status: Operating (subsidiary of Alphabet Inc.; converted from a Delaware corporation to a Delaware LLC effective 2017-09-30 — a form-of-entity change, not a distress event). Actively enforcing and defending patents in this technology area (Google LLC is both an IPR petitioner and, e.g., a district-court plaintiff against Sonos on family siblings — see signal 5 and signal 7).

---

Assignment timeline

Google Patents "Legal Events" for US 10,909,987 records two assignment-class events, both entered against this application. Important limitation: the retrieved record exposes the conveyance type, assignor, and assignee but not the reel/frame numbers or the correspondent of record for these two entries. I am therefore reporting the reel/frame and correspondent fields as not retrieved, rather than guessing. (Google Patents also does not display reel/frame; that data lives in USPTO Assignment Center, which I could not query programmatically for this number.)

• 2019-08-28 (recorded) — Reel/Frame: not retrieved

  • Conveyance: ASSIGNMENT OF ASSIGNORS INTEREST
  • Assignor: SHARIFI, MATTHEW (individual inventor)
  • Assignee: GOOGLE INC. (Delaware corporation)
  • Correspondent: not retrieved
  • Context: Initial inventor-to-corporate assignment — the standard take-title step, recorded alongside the filing of the 16/553,883 continuation; it re-papers the title from the inventor into the original assignee.

• 2019-08-28 (recorded) — Reel/Frame: not retrieved

  • Conveyance: ENTITY CONVERSION
  • Assignor: GOOGLE INC. (Delaware corporation)
  • Assignee: GOOGLE LLC (Delaware limited liability company)
  • Correspondent: not retrieved
  • Context: Change of name / internal reorg only — this papers Google Inc.'s conversion to Google LLC. The underlying Delaware Certificate of Conversion was filed 2017-09-29 and made effective 2017-09-30; the recording appears here dated to this application's 2019 filing. This is a form-of-entity conversion, not a sale, and conveys no change in beneficial ownership.

There are no other recorded assignments. No post-issuance transfer, no security agreement, no license recordation, no release, and no transfer to any third-party entity appears for this patent. On these facts the chain of title terminates at Google LLC, the original (converted) assignee, which still owns the patent.

Verification caveat: I could not open a populated USPTO Assignment Center abstract-of-title for 10,909,987 in this session; the two entries above are from Google Patents' Legal Events. Confirm the reel/frame numbers and correspondent by running the patent number at https://assignmentcenter.uspto.gov/ (or the indexed mirror https://assignment.uspto.gov/patent/index.html). If that search returns only these two Google-internal entries, the timeline is complete as shown; if it returns additional entries, they would supersede this reconstruction.

---

Timeline diagram

timeline
    title Ownership of US 10909987
    2014 : Provisional filed by Sharifi
    2015 : First non-provisional filed
         : Inventor assigns to Google Inc
    2017 : Google Inc converts to Google LLC
    2019 : Continuation filed
         : Chain recorded at USPTO
    2021 : Patent issues to Google LLC

---

NPE / troll-pattern signals

1. Shell-entity transfer — NOT PRESENT. The only downstream "transfer" is Google Inc. → Google LLC by ENTITY CONVERSION (effective 2017-09-30). No assignee carries an "IP / Patents / Licensing / Holdings / Ventures" suffix; the assignee is the operating parent itself, at its Mountain View HQ address, not a registered-agent service address. No single-purpose Delaware/Texas licensing LLC appears anywhere in the chain.

2. Known asserter in the chain — NOT PRESENT. Neither Google Inc. nor Google LLC matches any entry on the public NPE lists referenced (Acacia, Marathon, Intellectual Ventures, IPNav, Wi-LAN, Mosaid/Conversant, Vringo, Pendrell, Innovatio, MPHJ, Lumen View, Round Rock, Document Generation Corp, Spangenberg entities). Google LLC is the archetypal operating-company assignee, not a high-frequency-plaintiff NPE; it does not appear on Unified Patents / RPX high-frequency-plaintiff directories.

3. Repeat correspondent across the chain — NOT PRESENT / UNCLEAR. There is only one substantive assignor→assignee relationship here (inventor and name-change into the same corporate family), so there is no multi-link chain in which a single recording attorney could recur across shell LLCs. I could not retrieve the correspondent of record for either entry (see timeline limitation), so I flag this as unclear-but-moot rather than a null finding.

4. Cascading transfers — NOT PRESENT. There are two events, both dated 2019-08-28, both internal to the same corporate family (title cleanup + name change). There is no sequence of consecutive assignments through unrelated chained LLCs, and no shared-correspondent/common-principal pattern to evidence one.

5. Pre-litigation transfer — NOT PRESENT. No assignment is dated within six months of any infringement suit naming US 10,909,987 (none found — consistent with the prior litigation section). The only family litigation I could identify in this area is Google as plaintiff against Sonos on sibling patents (e.g., the N.D. Cal. action captioned Google LLC v. Sonos Inc., involving US 10,593,330), which is the opposite of an inbound NPE setup: the operating company is the one asserting, and it already owned the patents before filing.

6. Bankruptcy fire-sale — NOT PRESENT. No Chapter 7/11 proceeding, no sale-of-portfolio order, no distress transfer. Google Inc./Google LLC remained solvent and operating throughout; the entity conversion is a Delaware §18-214 conversion, not a liquidation.

7. Privateering — NOT PRESENT. Google did not transfer this patent to a separate NPE to assert on its behalf. The patent sits with Google LLC, and to the extent this technology has been enforced, it has been enforced directly by Google against a competitor (Sonos) on family siblings — i.e., first-party enforcement, not privateering.

8. Defensive aggregator (anti-NPE) — NOT PRESENT. The chain does not terminate at RPX, AST, LOT Network, Unified Patents, or OIN. It terminates at Google LLC. (Note the directional nuance: Google is a contributor/supporter of some defensive programs and a frequent IPR petitioner, but that is not the same as this patent having been acquired by a defensive aggregator.)

---

Verdict

Operating-company assertion.

Justification: The recorded chain is only two internal events — inventor Matthew Sharifi → Google Inc. (ASSIGNMENT OF ASSIGNORS INTEREST, recorded 2019-08-28) and Google Inc. → Google LLC (ENTITY CONVERSION, effective 2017-09-30, recorded 2019-08-28) — with no third-party transfer of any kind and no shell-LLC, security, or asserter entity anywhere in title. The current owner, Google LLC, is a large operating company that ships products embodying the claims (multi-device hotword arbitration across Android/Nest/Assistant hardware) and is itself the enforcing party in this technology area, having sued a competitor (Sonos) on family siblings. There are zero NPE signals present, so neither "NPE — high" nor "NPE — moderate" is tenable; the chain does not end at a defensive aggregator, so "Defensive / non-asserting" is also wrong. The one honest caveat: I confirmed no suit naming this exact number (US 10,909,987), so if you require a docketed assertion on the '987 patent itself to use the word "assertion," treat this as borderline with "Insufficient data" on the suing prong — but the ownership reconstruction is unambiguous and points to an operating-company-held patent.

Verify at: https://assignmentcenter.uspto.gov/ (search patent number 10909987) — and cross-check Google Patents Legal Events at https://patents.google.com/patent/US10909987/en .

---

Items I could not verify (stated explicitly rather than inferred)

• Reel/frame numbers for both recorded events — not retrievable in this session.
• Correspondent of record (attorney/firm) for both recorded events — not retrievable; the patent's "Attorney, Agent, or Firm" field was also not rendered in the text I retrieved.
• Exact recordation date vs. execution date for the entity-conversion entry: Google Patents shows 2019-08-28 against this application, while the underlying Delaware conversion was executed/effective 2017-09-30. I report both rather than collapsing them.

I'll search the USPTO and cross-reference the patent's citation records for US 10,909,987.

Let me retrieve the complete "References Cited" list for US 10,909,987 specifically.

I have partial data. Let me pin down the full citation list specifically attached to the '987 patent.

Prior-Art Analysis: US 10,909,987 B2 ("Hotword detection on multiple devices")

Note on method and a scope caveat up front. I anchored this to the reference list that the '987 patent itself carries in its (56) References Cited section, which I retrieved from the granted patent's PDF (patentimages.storage.googleapis.com/66/b6/bd/6fb1ff059b09fd/US10909987.pdf). That snippet rendered only part of the list. To fill the gaps I cross-checked against the (56) list of the sibling patent US 10,593,330 B2 (same family, same examiner, same specification) as reproduced in IPR2023-00119 Exhibit 1001 (Docket Alarm). The two lists share identical lead entries (4,363,102 Holmgren; 5,659,665 Whelpley) and identical foreign/NPL sections. I therefore treat the union below as the best available reconstruction of the '987 (56) list, but flag that I could not independently verify every U.S. patent entry against the '987 record itself. Where I could not verify a reference's subject matter, I say so rather than guess.

Also important: a reference appearing in a (56) list is not, by itself, anticipatory. These are references the examiner considered; the '987 claims issued over them. My §102 assessment below is a forward-looking "does this reference disclose the claim element" screen, not a statement that any of them was applied as an anticipation rejection.

---

A. The (56) References Cited for US 10,909,987

A1. U.S. Patent Documents

(The list continues past 8,938,394; the retrieve cut off. I did not obtain the tail.)

A2. U.S. Patent Application Publications

2014/0257821 A1 (Adams, G10L 25/93); 2014/0278383 A1 (Fan, G10L 25/84); 2014/0278435 A1 (Ganong, III, G10L 15/22); 2015/0081295 A1 (Yun); 2015/0154953 A1 (Bapat, G10L 15/06); 2015/0228274 A1 (Leppanen); 2015/0262577 A1 (Nomura, G10L 15/22); 2015/0379989 A1 (Balasubramanian); 2016/0104483 A1 (Foerster et al.); 2016/0155443 A1 (Khan); 2016/0189715 A1 (Nishikawa); 2016/0217790 A1 (Sharifi); 2016/0260431 A1 (Newendorp); 2016/0294973 A1 (Bakshi et al.); 2017/0083285 A1 (Meyers et al.).

A3. Foreign Patent Documents

JP S59-180599 (10/1984); JP H09-062293 (3/1997); JP H09-127973 (5/1997); JP H11-052976 (2/1999); JP H11-231896 (8/1999); JP 2000-075954 (3/2000); JP 2000-310999 (11/2000); JP 2006-227634 (8/2006); JP 2009-104020 (5/2009); JP 2014-092777 (5/2014); JP 2015-520409 (7/2015); JP 2019-021428 (2/2019); KR 10-2014-0031391 (3/2014); WO 1998/040875 (9/1998); WO 2013/006385 A1 (1/2013); WO 2013/163113 A1 (10/2013).

A4. Non-Patent Literature

• Auckenthaler et al., "Score Normalization for Text-independent Speaker Verification Systems," Digital Signal Processing, vol. 10 (2000), 13 pages.
• (The '330 counterpart carries additional NPL; the '987 record showed at least this item.)

---

B. §102 Anticipation Screen — the Most Relevant References

The '987 claims require, in combination, (i) local capture of an utterance, (ii) computation of a hotword-likelihood value, (iii) receipt of a second device's value, (iv) comparison, and (v) conditional activation of ASR/command execution, with dependents adding a threshold test, identifier exchange, timing windows, low-power mode, and short-range-radio transport. Only references that touch the multi-device score-exchange-and-arbitrate core can approach anticipation of independent claims 1, 7, and 13.

Bottom line on §102

• No single reference in the (56) list supplies all elements of independent claim 1, 7, or 13. Each missing element is the same one: receipt of a second computing device's hotword-likelihood value and comparison of it against the local device's value to gate activation. That is the point of novelty the '987 claims capture, and it is absent from every cited reference I could evaluate.
• The references most likely to have been applied (alone or in combination) in prosecution are the hotword/keyword-detection group — US 8,775,191 B1 (Sharifi) and US 2014/0278435 A1 (Ganong, III) — plus the multi-device coordination group (the Lambourne/Lindahl Sonos patents) and the score-normalization NPL (Auckenthaler). A §103 combination of the first two groups is the natural obviousness theory; §102 anticipation on any one reference is unlikely.

---

C. References that cannot be §102 prior art against these claims

• US 2016/0217790 A1 (Sharifi) — this is the applicant's own family member (publication of the priority-chain application). Same-family/same-inventor disclosures are not §102 prior art against the '987 claims; they are the disclosure from which the '987 claims descend.
• The '987's own parent chain (Ser. Nos. 16/171,495; 15/346,914; 15/088,477; 14/675,932; provisional 62/061,830) — priority documents, not art.
• Sibling members US 10,134,398, US 10,593,330, US 11,557,299, US 11,915,706, US 12,254,884 — same family.
• The Sonos patents (Lambourne/Lindahl) are true third-party art and can be §102/103 references; note the irony that Sonos later petitioned for IPR against the sibling US 10,593,330 (IPR2023-00119), per the earlier litigation section — consistent with the Lambourne/Lindahl citations being the family's most antagonistic art.

---

D. Confidence and caveats (read before relying on this)

1. Completeness: I reconstructed the (56) list from the '987 PDF (partial) plus the sibling '330 exhibit. The U.S. patent list is likely truncated after US 8,938,394 — several mid-2010s U.S. patents on the '330 record were not retrieved. Verify the full list at USPTO PatentCenter / the granted '987 PDF pages 1–2 before citing any item as definitively "cited in '987."
2. Descriptions: For roughly half the older U.S. patents (6,826,159; 6,973,426; 7,016,833; 7,222,072; 7,720,012; 8,099,288; 8,200,488; 8,209,174; 8,214,447; 8,340,975; 8,717,949; 8,719,009; 8,719,018; 8,805,890; 8,838,457) I could not verify subject matter from the retrieved sources. I have not invented descriptions for them. Their classifications and titles should be confirmed individually.
3. §102 mapping is a screen, not a rejection. The '987 claims issued over this art. My claim mapping is based on the independent-claim architecture (1 = method, 7 = system/low-power, 13 = CRM/ASR-before-comparison) described in the earlier summary section, whose exact per-claim wording carries the moderate-confidence caveat noted there.
4. Foreign/NPL: The JP/KR/WO numbers were retrieved verbatim; note the patent's own OCR renders some as "5S9-180599"/"H0141602" style — I have transcribed them literally per the rule not to auto-correct identifiers (JP S59-180599, JP H09-062293, etc.). Cross-check the ISO forms against J-PlatPat/KIPRIS/Espacenet.
5. One literal-identifier flag: the title of US 2014/0278435 A1 is sometimes listed as "Hotword detection" and its classification is G10L 15/22; I have not verified its full title/filing date beyond the listing. Do not treat the description as verbatim.

E. Recommended verification steps

1. USPTO PatentCenter → application 16/553,883 → "References Cited" / IFW → confirm the complete (56) list and any examiner-cited art.
2. Pull the '987 PDF pages 1–2 and transcribe the full (56) list directly.
3. For the hotword group, read US 8,775,191 B1 and US 2014/0278435 A1 against the '987 independent claims to confirm the single-device limitation (the key non-anticipation gap).
4. Check the family's prosecution histories — the '398/'330 files (which went through IPR2023-00119 and the June 9, 2026 CAFC reversal noted in the prior section) — for how the examiner and PTAB treated the same references, since the (56) lists are near-identical.

Obviousness Analysis — U.S. Patent No. 10,909,987 B2 ("Hotword detection on multiple devices")

Statute: 35 U.S.C. § 103 (AIA); KSR Int'l Co. v. Teleflex Inc., 550 U.S. 398 (2007); Graham v. John Deere Co., 383 U.S. 1 (1966).

---

0. Two threshold flags before the analysis begins

(a) Date inconsistency — flagged, not reconciled. The task header states "Current Date: April 26, 2026," but my searches returned a Federal Circuit decision dated June 9, 2026 (Google LLC v. Sonos, Inc., No. 24-2119) and commentary dated June 26, 2026. The system header for this session says 2026-10-01. Per the operating rule to prefer live search results over assumptions, I treat the June 9, 2026 CAFC decision as real ground truth and note the inconsistency rather than silently resolving it. Nothing in the §103 analysis below turns on this date discrepancy.

(b) Contradiction with the previously-generated claim summary. The prior "Patent summary" section describes claim 7 (system) as "tied to a low-power (sleep) mode posture." That is likely an artifact of conflating the '987 with its siblings US 10,134,398 and US 10,593,330, whose claims do recite "low power mode." The '987 specification text supplied in this task never uses the phrase "low power mode" — it uses "sleep state"/"awake state"/"activation state." The IPR record confirms this family-level fact: Petitioner's demonstratives state, of the '398/'330, "Common specification never mentions 'low power mode' as recited in the claims" (Docket Alarm, IPR2023-00118/-00119 Petitioner Demonstratives). I therefore analyze the '987's system claim both ways (with and without a low-power-mode limitation), because the answer materially differs. Confidence: high that the '987 spec lacks "low power mode"; moderate on whether '987 claim 7 imports it.

(c) Claim-text caveat carried forward. The granted claim set of the '987 was not fully rendered in the sources retrieved (Google Patents shows summary bullets; Justia fragments). The element numbers below are reconstructed labels, not verbatim claim text. Verify against the granted claims at USPTO PatentCenter or the patent PDF (patentimages.storage.googleapis.com/66/b6/bd/6fb1ff059b09fd/US10909987.pdf) before filing anything.

---

1. What constitutes the "prior art" here

The References Cited section on the face of the '987 (retrieved from the patent PDF) lists, inter alia:

Also relevant but not confirmed as being on the '987 face (it is the secondary reference in the sibling IPRs): US 2014/0163978 A1 (Basye), June 12, 2014 — per the CAFC opinion and the PTAB FWD, the Board held certain dependent claims of '398/'330 "unpatentable as obvious over Rosenberger in view of … Basye" (Justia CAFC No. 24-2119). I could not verify Basye's content in this session and do not assert what it discloses beyond its role in the IPR; treat the Basye-based grounds below as provisional.

Critical-date discipline. The '987 claims a provisional priority of Oct. 9, 2014 (Prov. 62/061,830). If the challenged claims are entitled to that date, then Rosenberger, Sharifi '191, Lambourne, Adams '821, Fan '383, Ganong '435 and Basye all pre-date it and are § 102(a)(1)/(a)(2) art. Several later-published face citations (US 2015/0081295 Yun; 2015/0154953 Bapat; 2015/0228274 Leppanen; 2015/0262577 Nomura; 2015/0379989 Balasubramanian; 2016/0104483 Foerster; 2016/0155443 Khan; 2016/0189715 Nishikawa; 2016/0217790 Sharifi; 2016/0260431 Newendorp; 2016/0294973 Bakshi; 2017/0083285 Meyers) are usable only if their effective filing dates precede Oct. 9, 2014 (§ 102(a)(2)). I have not verified those filing dates; a challenger must. Conversely, if Patent Owner obtains an earlier effective date via the § 119/§ 120 chain, or the challenger argues the claims are not supported by the 2014 provisional (a live risk given the "low power mode" terminology issue), the art set shifts. Confidence: moderate on the 2014 critical date applying to every '987 claim.

---

2. Person of ordinary skill in the art (POSITA)

Consistent with the IPR record (Petitioner's expert Dr. Michael T. Johnson: BS in CS Engineering and EE, MSEE, PhD in speech processing/ASR; a department chair in ECE), a POSITA here is reasonably defined as someone with a bachelor's degree in EE, CS, or a comparable field and ~2 years of experience in speech processing, ASR, or networked consumer-device software, or equivalent experience. That person would know: (i) hotword/keyword-spotting detectors producing confidence scores; (ii) always-on audio front ends (buffering, filtering, endpointing, digitizing); (iii) local-network discovery protocols (UPnP/SSDP — which the '987 spec itself concedes: "The communication protocol used between two computing devices may be universal plug and play"); and (iv) client/server ASR architectures.

---

3. Reconstructed claim scope and element-by-element mapping

3.1 Independent claim 1 (method) — core multi-device arbitration

3.2 Dependent-type features recited in the '987 summary of invention

• "determining that the first value satisfies a hotword score threshold" → Sharifi '191 (confidence-score-vs-threshold is the standard hotwordization primitive); Auckenthaler (2000) supplies normalized verification scores.
• "transmitting the first value to the second computing device … to a server, through a local network, or through a short range radio" → Rosenberger (wireless WS exchange); Lambourne ('014/'949: networked device coordination); UPnP, which the '987 spec itself names.
• "transmitting/receiving a first/second identifier" → device naming/registration (Lambourne; IP/MAC/name identifiers are routine).
• "determining an activation state … active / inactive" → Rosenberger's low-power "listening" mode vs. active mode.
• "particular amount of time has elapsed since receiving the audio data" → a latency-vs-certainty design choice; the PTAB record shows the family treated timing features as obvious over Rosenberger + Basye.
• "continuing, for a particular amount of time, to transmit the first value" → routine retransmission/reliability engineering; Rosenberger's repeated/broadcast signaling.
• "additional utterance / third device / inactive state" → mere repetition of the same steps.

3.3 The '987 claim variants that actually carry risk

• If claim 7 (system) is the "low power mode / messages exchanged while remaining in low power mode" variant (the '398/'330 formulation), then the June 9, 2026 CAFC decision is directly on point and narrows the obviousness/anticipation attack: the court held Rosenberger col. 8 "describes embodiments where upon detecting a spoken trigger or command phrase, the devices exit the low power 'listening' mode to calculate and exchange weighted signals," and therefore "fails to disclose the limitations directed to exchanging messages while in a low power mode." It also declined to affirm the alternative "different low-power mode" theory, remanding because the Board made no fact-findings.
• If claim 13 (medium) is the "begin ASR before the cross-device comparison resolves, then suppress" variant, the obviousness case must supply a latency-hiding motivation and must confront a teaching-away argument (Rosenberger's stated design principle is to avoid unnecessary activation/processing).

---

4. Grounds of rejection

I set out four grounds in descending order of strength.

Ground 1 — Claims 1–6: obvious over Rosenberger in view of Sharifi '191, further in view of Auckenthaler (2000)

Rationale.

1. Same field, same problem, analogous art. All three references concern voice-activated devices that must decide when a spoken trigger is directed at them. The '987's own Background concedes the problem was known ("it may be desirable to suppress responding to hotwords on devices that are not likely to be the ones a user intends to address").
2. Rosenberger supplies the entire multi-device architecture: multiple always-listening devices, each computing a device-specific weighting, exchanging that weighting, and self-selecting to respond. This maps to elements [1.1], [1.3], [1.4], [1.5].
3. Sharifi '191 supplies the missing numeric quantity. Rosenberger's "weighted signal" is already a measure of a device's relative suitability to handle the utterance; using an ASR-derived hotword confidence score as that weight is "a simple substitution of one known element for another," producing a predictable improvement (finer-grained discrimination between candidate devices) with no change in the principle of operation of Rosenberger's system. KSR, 550 U.S. at 416–17.
4. Auckenthaler provides the conventional normalization of verification scores to a common scale — needed only for the "normalized zero-to-one" flavor in the spec, and squarely enablement-level art for cross-device score comparability.
5. Reasonable expectation of success: high. Each step (hotwordize → score → transmit → compare → ASR) was independently known and mechanically composable.
6. No teaching away. Neither reference disparages combining a numeric confidence score with multi-device arbitration.

Ground 2 — Claims 7–12 (system): Ground 1 + Lambourne ('014/'949) + Basye, with the low-power question contested

• Without a "low power mode" limitation in the '987 claim, Ground 2 collapses into Ground 1 with the device-identification/registration and timing features supplied by Lambourne (networked device coordination/registration) and Basye (the PTAB relied on Basye for timing/related dependent features in the sibling IPRs). Motivation to combine: a device cannot exchange arbitration messages without first knowing which peers are hotword-responsive and how to address them — a straightforward engineering prerequisite, and the '987 spec itself concedes the answer ("may be universal plug and play").
• With a "low power mode" limitation, the § 103 case is materially harder. The § 103 argument would be: (i) Rosenberger teaches operating devices in a low-power listening mode; (ii) it would have been obvious to keep the devices in that low-power mode during the WS exchange to capture the energy savings that motivated the low-power mode in the first place; (iii) hence the combination. But this is precisely the theory the Federal Circuit rejected as an anticipation rationale on June 9, 2026, and remanded on the alternative low-power-mode theory for lack of Board fact-findings. A § 103 framing is legally distinct (no single-reference requirement) and could succeed where anticipation failed — but a challenger would need evidence that a POSITA would have regarded message exchange while remaining in low power as the natural design, not merely as a possible one. Confidence: low-to-moderate absent expert testimony.

Ground 3 — Claims 13–17 (medium, "start ASR early, suppress later"): Ground 1 + Ganong '435 (and/or Adams '821 / Fan '383)

Rationale. Ganong '435 (published Sep. 18, 2014, on the '987 face) discloses network-connected voice-controlled devices that begin processing after a wakeword; Adams '821 and Fan '383 address operating-state/noise conditions. The § 103 theory is:

1. Shifting the start of ASR to the moment a device's own hotword score crosses threshold — while the cross-device comparison is still pending — is a known buffering/latency-hiding technique, and Rosenberger already contemplates a device continuing to function before final arbitration.
2. The predictable result is faster responsiveness if the device wins, and no user-visible effect if it loses (the '987 spec confirms: the losing device "does not appear, to the user, to be active").
3. Motivating combination: minimizing apparent latency after the hotword, a recognized design objective in voice interfaces.

Counter-risk (must be briefed). This ground most invites a teaching-away rebuttal: Rosenberger's premise is avoiding work on non-selected devices, so proactively running ASR on every device arguably cuts against its principle of operation — an argument with particular force for a "low power mode" claim. Expect Patent Owner to press In re Gordon-style reasoning and to emphasize that the discarded-ASR energy cost is exactly what the asserted invention purports to avoid. Confidence: moderate for a non-low-power medium claim; low if the medium claim recites low-power operation.

Ground 4 — Single-reference fallback: Rosenberger alone anticipates (and therefore renders obvious) claim 1

Because the core limitations [1.1]–[1.5] map so closely onto Rosenberger's weighted-signal exchange and device self-selection, a § 102 case on claim 1 is plausible, and a § 103 case follows a fortiori. Note, however, that the Board's anticipation finding on this theory was reversed on June 9, 2026 for the sibling claims (on the low-power-mode limitation, not on the score-exchange concept). If '987 claim 1 contains no low-power limitation, that reversal is largely inapposite, and the Sonos IPR record remains a usable template.

---

5. Motivation-to-combine statement (for a KSR-compliant office action)

A POSITA at October 2014, facing the known problem of multiple always-listening devices simultaneously triggering on one hotword, would have been motivated to look to Rosenberger, which teaches arbitrating among plural voice-responsive devices via exchanged weighted signals, and to Sharifi '191, which teaches computing a normalized hotword confidence score by extracting audio features and comparing the score to a threshold. The POSITA would have been further motivated to substitute Sharifi's confidence score for Rosenberger's generic weight because a numeric likelihood is a strictly more informative basis for the very comparison Rosenberger performs, and would have been motivated to add a threshold test before transmission to reduce needless network traffic and to avoid waking non-target devices. The combination arranges known elements in a known way to achieve the predictable result of selecting one device to perform speech recognition — i.e., the "improvement is not more than the predictable use of prior art elements according to their established functions" (KSR).

---

6. Rebuttals and weaknesses a challenger must address

---

7. Bottom line

• Strongest § 103 case: Rosenberger + Sharifi '191 (+ Auckenthaler) against the '987's independent method claim and its threshold/normalization/transmission/identifier/activation-state dependents. Every element is accounted for, the references are analogous art in a shared field, the substitution of a hotword confidence score for Rosenberger's weighted signal is a classic KSR substitution with predictable results, and Sonos's IPR2023-00118/-00119 record supplies a tested factual template for this reference pairing (with Basye as the timing/dependent-claim secondary).
• Contested variants: Any '987 claim reciting operation "while remaining in low power mode" is materially harder — that is the exact limitation the Federal Circuit found Rosenberger did not disclose in its June 9, 2026 decision (Google LLC v. Sonos, Inc., No. 24-2119), reversing and remanding the PTAB's anticipation holdings on US 10,134,398 and US 10,593,330. That decision is about sibling patents and an anticipation theory; it does not by itself immunize the '987, and it expressly leaves open the alternative "different low power mode" theory on remand.
• Weakest variant: the medium claim's "begin ASR before the cross-device comparison resolves, then suppress" behavior, where a teaching-away defense is colorable.

Given the as-delivered sources, I could not verify (i) the verbatim granted claim text of the '987, (ii) the exact content of US 2014/0163978 (Basye), or (iii) the effective filing dates of the post-October-2014 face citations. Verify all three before relying on this analysis for any filing or challenge.

---

Sources consulted

• Google Patents, US 10,909,987 B2 — https://patents.google.com/patent/US10909987/en
• US 10,909,987 B2 patent PDF (References Cited; 17 claims) — https://patentimages.storage.googleapis.com/66/b6/bd/6fb1ff059b09fd/US10909987.pdf
• Family member EP 3 139 378 B1 (cites US 8,775,191 B1 as [HOTWORD][QUERY] background) — http://data.epo.org/pise-server/rest/collections/lgpi/EP3139378B1.pdf
• Google LLC v. Sonos, Inc., No. 24-2119 (Fed. Cir. June 9, 2026) (nonprecedential) — https://law.justia.com/cases/federal/appellate-courts/cafc/24-2119/24-2119-2026-06-09.html ; https://www.courtlistener.com/opinion/[10872385](/patent/10872385)/google-llc-v-sonos-inc/
• IPWatchdog, "Federal Circuit Reverses PTAB Findings of Unpatentability for Google's 'Hotword' Patents" (June 9, 2026) — https://ipwatchdog.com/2026/06/09/federal-circuit-reverses-ptab-findings-of-unpatentability-for-googles-hotword-patents/
• Fitch, Even, Tabin & Flannery, "Federal Circuit Reverses PTAB Anticipation Ruling and Revives Google 'Hotword' Patents" (June 26, 2026) — https://www.fitcheven.com/2026/06/26/federal-circuit-reverses-ptab-anticipation-ruling-and-revives-google-hotword-patents/
• IPR2023-00118/-00119, Petitioner Demonstratives (Rosenberger, col. 8; "Common specification never mentions 'low power mode'") — https://www.docketalarm.com/cases/PTAB/IPR2023-00119/Sonos_Inc._v._Google_LLC/docs/02-29-2024-Petitioner/Exhibit-1034-Petitioners_Demonstratives.pdf
• Exhibit 1001, US 10,593,330 (Rosenberger cited art list) — https://www.docketalarm.com/cases/PTAB/IPR2023-00119/Sonos_Inc/docs/10-27-2022-Petitioner/Exhibit-1001-US10593330.pdf
• Exhibit 2012, Declaration of Stuart Lipoff (IPR2023-00119) — https://www.docketalarm.com/cases/PTAB/IPR2023-00119/Sonos_Inc._v._Google_LLC/docs/08-11-2023-Patent_Owner/Exhibit-2012-Declaration_of_Stuart_J_Lipoff.pdf
• Google LLC v. Sonos, Inc., N.D. Cal. No. 3:22-cv-04552, Complaint (cited-art listing) — https://www.docketalarm.com/cases/California_Northern_District_Court/3--22-cv-04552/Google_LLC_v._Sonos_Inc/docs/1/1.pdf
• RPX Insight, US 10,593,330 B2 claim 1 (family claim structure) — https://insight.rpxcorp.com/patent/[US10593330B2](/patent/US10593330B2)