Sanas.AI Inc. v. Krisp Technologies, Inc.

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
United States District Court
Northern District of California
UNITED STATES DISTRICT COURT
NORTHERN DISTRICT OF CALIFORNIA
SANAS.AI INC.,
Plaintiff,
v.
KRISP TECHNOLOGIES, INC.,
Defendant.
Case No. 25-cv-05666-RS
ORDER DENYING JUDGMENT ON
THE PLEADINGS
I. INTRODUCTION
Defendant Krisp Technologies (“Krisp”) brings the present motion for judgment on the
pleadings to dismiss various patent infringement claims brought by Plaintiff Sansa AI (“Sanas”)
related to their patented accent conversion technology. Unpersuaded by Krisp’s representation of
each patent as directed towards longstanding human activities, patent ineligible format conversion,
generic computer functions, results rather than methods, or some combination thereof, the motion
is denied as to each count as set forth below.1
II. BACKGROUND2
Having discussed the facts of this suit in this court’s prior orders, see e.g., Dkt. 58, only
those facts relevant to this motion are included here. Sanas and Krisp are competitors in the
1 Pursuant to Civil Local Rule 7-1(b), the motion is suitable for disposition without oral argument,
and the hearing set for February 26, 2026 is vacated.
2 Unless otherwise stated, this order accepts well-pled factual allegations made in the complaint
and counterclaims as true.
Case 3:25-cv-05666-RS Document 72 Filed 02/23/26 Page 1 of 14

ORDER DENYING JUDGEMENT ON THE PLEADINGS
CASE NO. 25-cv-05666-RS
2
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
United States District Court
Northern District of California
market for real-time audio-enhancing technology.
Krisp was founded as 2Hz in 2018 and entered the market first with background noise
suppression software. Concurrent with the release of its noise suppression technology, Krisp filed
a provisional patent application that was awarded on August 24, 2021 as U.S. Patent No.
11,100,941 (“the ’941 Patent”) for Speech Enhancement and Noise Suppression Systems and
Methods.
Meanwhile, the Sanas founders were focused on accent conversion. Sanas was founded by
Maxim Serebryakov, Shawn Zhang, and Andrés Pérez Soderi, three former Stanford classmates,
along with Sharath Keshava Narayana, an entrepreneur. Between 2020 and 2021, the Sanas
founders built out a database of accented human speech, including identifying and recording
“ideal” target accents. They then developed software that translates accents between parties in
real-time, fast enough and at a high enough quality that normal conversation can be carried out,
leveraging unique student-teacher machine learning architecture and parallel data generation.
Sanas launched their first product in 2021 and filed a provisional patent application,
resulting in U.S. Patent No. 11,948,550 (“the ’550 Patent”), Real-time Accent Conversion Model,
on April 2, 2021. Other applications followed, resulting in five more patent awards by September
2025: U.S. Patent No. 12,125,496 (“the ’496 Patent”), Methods for Neural Network-Based Voice
Enhancement and Systems Thereof, issued on October 22, 2024; U.S. Patent No. 12,131,745 (“the
’745 Patent”), System and Method for Automatic Alignment of Phonetic Content For real-Time
Accent Conversion, issued on October 29, 2024; U.S. Patent No. 11,715,457 (“the ’457 Patent”),
Real Time Correction of Accent in Speech Audio Signals, issued on August 1, 2023; U.S. Patent
No. 12,412,561 (“the ’561 Patent”), Real Time Correction of Accent in Speech Audio Signals,
issued on September 9, 2025; and U.S. Patent No. 12,417,756 (“the ’756 Patent”), Systems and
Methods for Real-Time Accent Mimicking, issued on September 16, 2025. Sanas is the owner of
record of each of the asserted patents and owns all rights in each of them, including without
limitation all rights to recover for past infringement thereof. Collectively, these are the “Sanas
Accent Conversion Patents.”
Case 3:25-cv-05666-RS Document 72 Filed 02/23/26 Page 2 of 14

ORDER DENYING JUDGEMENT ON THE PLEADINGS
CASE NO. 25-cv-05666-RS
3
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
United States District Court
Northern District of California
As discussed in this court’s prior order, Dkt. 58, Krisp and Sanas engaged in extensive
discussions about a partnership between July 2021 and November 2022, during which details
about Sanas’ accent conversion technology were shared with Krisp. Krisp abruptly terminated the
discussions in November 2022, and no collaboration ensued.
On April 27, 2023, sixteen months after Krisp first reached out to Sanas about licensing its
technology, Krisp publicly announced “early access” to its proprietary Krisp AI Accent
Conversion technology. Three months later, on July 21, 2023, Krisp filed two provisional U.S.
patent applications, which issued as U.S. Patent No. 12,205,609 (“the ’609 Patent”) and U.S.
Patent No. 12,223,979 (“the ’979 Patent”) (collectively, the “Krisp Accent Conversion Patents”).
The Krisp Accent Conversion Patents cite Sanas’ parent applications for the ’550 and ’745
Patents. They also cite Sanas’ asserted ’496 Patent and ’457 Patent, which shares a parent
application with the ’561 Patent.
Sanas alleges that the Krisp Accent Conversion Patents do not disclose or claim any novel
or non-obvious approaches to generating parallel data for speech conversion but rather claim as
Krisp’s own inventions techniques and approaches discussed in Sanas’ earlier patent applications
and issued patents. Sanas also alleges that the Krisp Accent Conversion Patents rely and build
upon information provided by Sanas employees, including co-founder Serebryakov, such that the
Sanas employees should have been named as joint inventors. In March 2025, Krisp launched
“Krisp AI Accent Conversion v3.” Krisp also markets a product called “Accent Localization.”
This lawsuit followed. Sanas’ first amended complaint, filed in September 2025, asserts
patent claims, among other claims against Krisp, for alleged willful infringement of the ’550,
’496, ’745, and ’457 Patents (counts one, two, three, and four, respectively) and infringement of
the ’561 and ’756 Patents (counts eight and nine, respectively). Krisp brought counterclaims
including one for patent infringement of Krisp’s asserted ‘941 Patent (count one) for making,
using, offering for sale, and selling Sanas’ noise cancellation products.
In September and October 2025, Krisp and Sanas moved to dismiss each other’s non-
infringement claims and counterclaims under Rule 12(b)(6). In November 2025, Krisp made a
successive motion to dismiss Sanas’ patent infringement claims. On December 1, 2025, this court
Case 3:25-cv-05666-RS Document 72 Filed 02/23/26 Page 3 of 14

ORDER DENYING JUDGEMENT ON THE PLEADINGS
CASE NO. 25-cv-05666-RS
4
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
United States District Court
Northern District of California
denied both parties’ motions to dismiss the non-infringement claims and counterclaims in their
entirety except for Sanas’ motion to dismiss count three of Krisp’s counterclaims for violation of
California Business and Professions Code § 17044. Dkt. 58. On December 19, 2025, this court
also denied Krisp’s successive motion to dismiss Sanas’ claims for patent infringement on
procedural grounds under Rule 12(g)(2). Dkt. 63.
With the pleadings now closed, Krisp makes the present motion for judgement on the
pleadings, renewing its arguments from its second motion to dismiss that Sanas’ patents reflect
non-patentable abstract ideas. Specifically, Krisp argues that counts one, two, and nine of Sanas’
first amended complaint should be dismissed because the ’550, ’496, and ’756 Patents are directed
to the abstract idea of translation, converting audio data from one form to another, and counts four
and eight should be dismissed because the ’457 and ’561 Patents are directed to the abstract idea
of gathering data, analyzing it, and displaying results.
III. LEGAL STANDARD
Patent eligibility “is a question of law, based on underlying facts” that “may be… resolved
on a [Rule 12(c)] motion,” see SAP Am., Inc. v. InvestPic, LLC, 898 F.3d 1161, 1166 (Fed. Cir.
2018), which is “substantially identical” to the analysis under a Rule 12(b)(6) motion, see Chavez
v. United States, 683 F.3d 1102, 1108 (9th Cir. 2012). The primary difference is procedural: Rule
12(c) permits a party to move for judgment on the pleadings after the pleadings close but “early
enough not to delay trial[.]” Fed. R. Civ. P. 12(c). All allegations must be accepted as true and
construed in the light most favorable to the non-moving party: in this case, Sanas. See Turner v.
Cook, 362 F.3d 1219, 1225 (9th Cir. 2004). Furthermore, the moving party, i.e., Krisp, bears the
burden of demonstrating invalidity by clear and convincing evidence. See 35 U.S.C. §
282; Microsoft Corp. v. I4I Ltd. P'ship, 564 U.S. 91, 95, 131 S.Ct. 2238, 180 L.Ed.2d 131 (2011).
IV. DISCUSSION
Under Section 101 of the Patent Act, “[w]hoever invents or discovers any new and useful
process, machine, manufacture, or composition of matter, or any new and useful improvement
thereof, may obtain a patent therefore,” 35 U.S.C. § 101, but the Supreme Court has “long held
Case 3:25-cv-05666-RS Document 72 Filed 02/23/26 Page 4 of 14

ORDER DENYING JUDGEMENT ON THE PLEADINGS
CASE NO. 25-cv-05666-RS
5
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
United States District Court
Northern District of California
that this provision contains an important implicit exception: Laws of nature, natural phenomena,
and abstract ideas are not patentable.” Alice Corp. Pty. v. CLS Bank Int’l, 573 U.S. 208, 216, 134
S. Ct. 2347, 2354, 189 L. Ed. 2d 296 (2014) (internal quotation marks and citations omitted).
“[S]uch discoveries are manifestations of… nature, free to all men and reserved exclusively to
none.” Mayo Collaborative Servs. v. Prometheus Labs., Inc., 132 S. Ct. 1289, 1293
(2012) (internal quotation marks and citations omitted). On the other hand, “[a]t some level, all
inventions… embody, use, reflect, rest upon, or apply laws of nature, natural phenomena, or
abstract ideas,” and so “[a]pplications of such concepts to a new and useful end… remain eligible
for patent protection.” Alice, 573 U.S. at 217 (internal quotation marks and citations omitted).
“The Supreme Court has established a two-step framework,” the Alice test, “for ‘distinguishing
patents that claim laws of nature, natural phenomena, and abstract ideas from those that claim
patent-eligible applications of those concepts.’ ” Trading Techs. Int'l, Inc. v. IBG LLC, 921 F.3d
1084, 1092 (Fed. Cir. 2019) (citing Alice, 573 U.S. at 217).
The first Alice step asks, “whether the claims at issue are directed to one of those patent-
ineligible concepts.” Alice, 573 U.S. at 217. The inquiry does not focus on the results achieved by
the patents, but instead “must turn to any requirements for how the desired result is achieved.”
Elec. Power Grp., LLC v. Alstom S.A., 830 F.3d 1350, 1355 (Fed. Cir. 2016) (emphasis in
original). See also McRO, Inc. v. Bandai Namco Games Am. Inc., 837 F.3d 1299, 1314 (Fed. Cir.
2016) (Alice step one inquiry “look[s] to whether the claims… focus on a specific means or
method that improves the relevant technology or are instead directed to a result or effect that itself
is the abstract idea and merely invoke generic processes and machinery.”).
For computer-related claims, courts look to whether the claims “improve the functioning of
the computer itself,” Alice, 573 U.S. at 219, or whether “computers are invoked merely as a tool”
to implement an abstract process, Enfish, 822 F.3d at 1336. “[A]dding a computer cannot spare a
claim that otherwise would be directed to an abstract idea[.]” Visual Memory LLC v. NVIDIA
Corp., 867 F.3d 1253, 1260 (Fed. Cir. 2017). Claims that “delineate steps through which the
machine learning technology achieves an improvement” or otherwise “disclos[e] improvements to
Case 3:25-cv-05666-RS Document 72 Filed 02/23/26 Page 5 of 14

ORDER DENYING JUDGEMENT ON THE PLEADINGS
CASE NO. 25-cv-05666-RS
6
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
United States District Court
Northern District of California
the machine learning models to be applied” are likely to be directed to a patent eligible invention.
See Recentive Analytics, Inc. v. Fox Corp., 134 F.4th 1205, 1213, 1215-16 (Fed. Cir. 2025).
However, claims which disclose only “the application of generic machine learning to new data
environments” are not. Id.
If the claims are directed at a patent-ineligible concept, the second Alice step asks whether
“the elements of each claim both individually and as an ordered combination… transform the
nature of the claim into a patent-eligible application.” Alice, 573 U.S. at 216. This step entails the
“search for an inventive concept—i.e., an element or combination of elements that is sufficient to
ensure that the patent in practice amounts to significantly more than a patent upon the [ineligible
concept] itself.” Id. at 217–18 (internal citations and quotations omitted). “For the role of
a computer in a computer-implemented invention to be deemed meaningful in the context of this
analysis, it must involve more than performance of well-understood, routine, [and] conventional
activities previously known to the industry.” Content Extraction & Transmission LLC v. Wells
Fargo Bank, N.A., 776 F.3d 1343, 1347–48 (Fed. Cir. 2014). “[T]he mere recitation of a generic
computer cannot transform a patent-ineligible abstract idea into a patent-eligible invention.” Id. at
1348. However, “an inventive concept can be found in the non-conventional and non-generic
arrangement of known, conventional pieces.” BASCOM Glob. Internet Servs., Inc. v. AT & T
Mobility LLC, 827 F.3d 1341, 1350 (Fed. Cir. 2016).
Krisp argues the ’550, ’496, and ’756 Patents are directed to the abstract idea of
translation, converting audio data from one form to another, and thus patent ineligible and the ’457
and ’561 Patents are directed to the abstract idea of gathering, analyzing, and displaying certain
results.
A. The ’550 Patent
The ’550 Patent, Real-time Accent Conversion Model, achieves specific improvements
over other accent conversion technology, but, despite Krisp’s insistence otherwise, the patent is
not directed at those achievements. Instead, the ’550 Patent is directed at particular methods,
which achieve those improvements. See Elec. Power Grp., 830 F.3d at 1355 (The inquiry “must
Case 3:25-cv-05666-RS Document 72 Filed 02/23/26 Page 6 of 14

ORDER DENYING JUDGEMENT ON THE PLEADINGS
CASE NO. 25-cv-05666-RS
7
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
United States District Court
Northern District of California
turn to any requirements for how the desired result is achieved.”). Nor does the ’550 Patent reflect
longstanding human activities or digital format conversation such that it is not patentable.
The ’550 Patent improves upon prior accent conversion strategies that “adjust the audio
characteristics ( e.g., pitch, intonation, melody, stress) of a first speaker’s voice” and others that
“involve[] a speech-to-text (STT) conversion of input speech as a midpoint, followed by a text-to-
speech (TIS) conversion to generate the output[.]” Dkt. 35-1 at 1:44-51, 60-63. The former “type
of approach does not account for the different pronunciations of certain sounds that are inherent to
a given accent” such that if a speaker pronounces “th” as “d,” the output will do the same. Id. at
1:44-51. The latter approach does not “capture many of the nuances of input speech” and
“generally involves a degree of latency (e.g., up to several seconds) that makes it impractical for
use in real-time communication[.]” Dkt. 35-1 at 1:44-51; id. at 1:64-2:2. The ’550 Patent is
directed at methods that improve upon these: identifying non-text linguistic units of speech
common to an accent, mapping those to other non-text linguistic units of speech common to a
second accent, and, finally, generating an audio output of the first speech in the accent of the
second.
Krisp conclusively explains that this technology automates “longstanding human
activities,” which is a telltale sign of abstraction, Dkt. 40 at 6, but translating non-linguistic sounds
is not a longstanding human activity. Krisp also conclusively likens Sanas’ accent conversion to
format conversions which are considered abstract ideas. However, the authorities Krisp’s cites for
this proposition all relate to the conversion of data from one digital format to another. See e.g.,
Hawk Tech. Sys., LLC v. Castle Retail, LLC, 60 F.4th 1349, 1357 (Fed. Cir. 2023) (digitizing
images and converting video images into selected video formats with particular resolution);
Adaptive Streaming, Inc., v. Netflix, Inc., 836 F.App’x 900, 901 (Fed. Cir. 2020) (unpublished)
(converting incoming videos in one format from one device type to a second format better-suited
for a second device); GoTV Streaming, LLC v. Netflix, Inc., No. 2024-1669, 2026 WL 346200, at
*9 (Fed. Cir. Feb. 9, 2026) (“using a template for tailoring the final set of specifications the device
uses to put an image on the screen, with use of the ordinary capacity of computers (including a
Case 3:25-cv-05666-RS Document 72 Filed 02/23/26 Page 7 of 14

ORDER DENYING JUDGEMENT ON THE PLEADINGS
CASE NO. 25-cv-05666-RS
8
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
United States District Court
Northern District of California
given wireless device) and networks to aid (in automating or speeding up, e.g.) that process”). In
contrast, Sanas’ technology translates between audio “formats” whose differences are intelligible
and meaningful to humans apart from the devices rendering them, making the cited authority
unconvincing.
Even if Sanas’ conversion of speech in one accent to that of another were an abstract idea
as Krisp contends, Krisp’s argument that the claims are not directed at particular methods because
the claim steps do not “recit[e] any specific technical mechanism” for the conversion is plainly
untrue. Dkt. 40 at 5. The steps of claim 1 include: via one machine-learning algorithm, “aligning
and classifying…frames of… captured speech content” of a first accent such that the frames
“correspond[] to respective [frames] of [a] plurality of different speakers;” via that same
algorithm, “deriv[ing] a non-text linguistic representation of the set of phonemes” of received
speech having first accent; via a second machine-learning algorithm, “mapping… [a] non-text
linguistic representation of a…phoneme… associated with the first [accent] to a second non-text
linguistic representation of a second phoneme…associated with a second” accent; “convert[ing]”
synthesized audio representative of that map “into a synthesized version of” inputted speech
“having the second accent” which “comprises the updated set of phonemes[.]” Dkt. 35-1 at 10.
Despite Krisp’s best efforts to excerpt these steps and reduce them to their most generic
data-processing terms,3 the claim steps reflect more than “generic computing and machine-
learning technology” “incorporated in purely functional terms.” Dkt 40 at 15. They reflect frame
alignment, identification of and mapping across non-text linguistic representations, and frame wise
mapping—alleged innovations which allow speech to be outputted in a second accent while
retaining key vocal characteristics in near real-time. In Enfish as discussed in Visual Memory, “the
district court’s characterization of the claims as being ‘directed to the abstract idea of “storing,
3 Krisp’s excerpt emphasizes the verbs “training… applying… synthesizing… converting,”
reducing the patent to those generic terms, makes only cursory reference to the non-text linguistics
representations, and omits reference to the frames and phonemes. Dkt. 40 at 9 (emphases added in
Krisp’s filing). Adopting Krisp’s characterization, “describing the claims at a high level of
abstraction, divorced from the claim language itself” and “overgeneralizing [the] claims,”
however, would be reversible error. See, e.g., Contour IP Holding LLC v. GoPro, Inc., 113 F.4th
1373, 1379–80 (Fed. Cir. 2024); TecSec, Inc. v. Adobe Inc., 978 F.3d 1278, 1293 (Fed. Cir. 2020).
Case 3:25-cv-05666-RS Document 72 Filed 02/23/26 Page 8 of 14

ORDER DENYING JUDGEMENT ON THE PLEADINGS
CASE NO. 25-cv-05666-RS
9
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
United States District Court
Northern District of California
organizing, and retrieving memory in a logical table” ’ ” was “rejected” because the claims
“recit[ed] a self-referential table.” Visual Memory LLC v. NVIDIA Corp., 867 F.3d 1253, 1258
(Fed. Cir. 2017) (discussing and citing Enfish, LLC v. Microsoft Corp., 822 F.3d 1327, 1333, 1337
(Fed. Cir. 2016)). “The [claim] specification described the benefits of using a self-referential
table—faster searching and more effective data storage—and highlighted the differences between
the claimed self-referential table and a conventional database structure.” Id. (citing Enfish, 822 F.
3d at 1336). Here, the identification of frames and non-linguistic units of sound are akin to
reference to self-referential tables in Enfish.
In GoTV Streaming, LLC v. Netflix, Inc., for which Krisp filed a statement of recent
decision, Dkt. 69, the Federal Circuit considered the patent eligibility of claims related to tailoring
video content to specific mobile device screens. The claims “refer[red] to the ‘algorithm’ and
‘architecture’ ”— “highly general labels”— “without further identification of details,” and were
held to be not patent eligible. GoTV Streaming, LLC v. Netflix, Inc., No. 2024-1669, 2026 WL
346200, at *9 (Fed. Cir. Feb. 9, 2026). The Court indicated the outcome would have been different
had the claims referred to specific “ ‘structure’ (akin to header/payload or bit-slot assignments) of
the bit collections transferred between server and wireless device” that improved tailoring video
content to specific mobile device screens. Id. As if heeding this advice, the ’550 Patent claims do
not rely on highly general labels but rather invoke specific data segments such as frames and
phonemes and processes such as non-linguistic mapping.
In sum, the ’550 Patent claims are not directed to an abstract idea because they are directed
at specific technical mechanisms, leveraging non-linguistic frame alignment/classification and
mapping, for accent conversion. Accordingly, we need not reach the second Alice step.4
4 As many cases have noted, though, “the two [Alice] stages are plainly related…[T]he two stages
involve overlapping scrutiny of the content of the claims, but... there can be close questions about
when the inquiry should proceed from the first stage to the second.” Elec. Power Grp., 830 F.3d at
1353 (citations omitted). In other words, the step one analysis here reflects how a step two
analysis, asking whether “an inventive concept can be found in the non-conventional and non-
generic arrangement of known, conventional pieces,” BASCOM, 827 F.3d at 1350, likely would
result.
Case 3:25-cv-05666-RS Document 72 Filed 02/23/26 Page 9 of 14

ORDER DENYING JUDGEMENT ON THE PLEADINGS
CASE NO. 25-cv-05666-RS
10
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
United States District Court
Northern District of California
B. ’756 Patent
Similar to the ’550 Patent, the ’756 Patent, Systems and Methods for Real-Time Accent
Mimicking, is directed at non-abstract methodological improvements. The ’756 Patent extracts
features from voices, identifies those associated with the accent and those associated with a
particular speaker’s natural voice, combines them, and translates audio in one accent to audio in
another, with the natural vocal characteristics of the speaker preserved. The alleged claim
advances include extraction of specific features (e.g., pitch, speed, intonation, phoneme
pronunciation, prosodic features, articulation, timbre, rhythm); classification and identification of
those features between those associated with an accent (e.g., pitch, speed. intonation, phoneme
pronunciation) and those associated with a speaker’s unique vocal characteristics (e.g., pitch,
speed, prosodic features, intonation, articulation, timbre, rhythm) 5 along with specifications about
how that is carried out (e.g., mel frequency cepstral coefficient fingerprinting, identity encodings);
development of a voice “fingerprint” reflecting the natural sound of a speaker via “mel frequency
cepstral coefficient (MFCC) analysis”; and synthesis of the accent features and voice fingerprint to
generate audio that mimics the speaker in a secondary accent via a specified pipeline process with
steps invoking these processes (extraction, classification, identification, recombination) along the
way. 6 Dkt. 36-6 at 9-10.
Krisp recycles much of its arguments for the ’550 Patent here: simplifying the ’756 claim 1
steps into generic terms, characterizing the ’756 Patent as directed at “a longstanding human
5 Note, an accent has unique intonation as does a speaker’s vocal style. This process would
identify that associated with the accent and that associated with the vocal style.
6 In the specific language of the patent, the claim 1 steps include: “extract[ing] accent features of
first input speech associated with a first accent of a first user;” “generat[ing] characteristics
specific to a natural voice of [a] second user” with a second accent “wherein the generated
characteristics correspond to vocal traits that are distinct to the second user and comprise one or
more of a voice quality or one or more phonetic patterns, prosodic features, articulation styles, or
intonation patterns[;]” modifying second input speech of a second user and synthesizing it “based
on the generated characteristics and the extracted accent features” to “mimic[] the first accent”
while “preserv[ing] aspects of the natural voice of the second user”; and “provid[ing]… output
audio… generated based on the modified version of the second input speech.” Id.
Case 3:25-cv-05666-RS Document 72 Filed 02/23/26 Page 10 of 14

ORDER DENYING JUDGEMENT ON THE PLEADINGS
CASE NO. 25-cv-05666-RS
11
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
United States District Court
Northern District of California
activity,” Dkt. 40 at 14, and comparing the ’756 Patent’s accent conversion to digital format
conversions that have failed the first Alice step. To be sure, humans have been mimicking accents
for a long time. Yet, this is not dispositive. Krisp’s other arguments fail for the same reasons as
they did before: the ’756 Patent is directed at particular processes related to selective feature
preservation and recombination that improve accent mimicking, and which are not analogous to
reformatting for enhanced digital performance.
C. ’496 Patent
The’496 Patent, Methods for Neural Network-Based Voice Enhancement and Systems,
Thereof, identifies and suppresses non-linguistic and non-meaningful noise in speech audio to
generate clearer speech while “enhancing” the speaker’s natural voice. The ’496 Patent does this
through a series of steps: “fragment[ing] input audio data into… input speech frames…
compris[ing] foreground speech content, … non-content elements, and … speech characteristics;”
using “a first neural network,” “convert[ing] the input speech frames to low-dimensional
representations” “omit[ing]” some non-content elements; using “a second neural network,”
“generat[ing] target speech frames” from the low-dimensional representations of the input speech.
Dkt. 35-2 at 9:60-10:11. Claims 4 and 5 detail how the first neural network is to be trained
including augmentations, a process of artificially increasing the size and diversity of training data
without collecting more data, for features like background noise, masked data, microphone pops,
and convolving speech. Id. 10:39-43. Claims 6 and 7 detail how the second neural network will be
trained and structured, including specifying what models should be used for this network (a
combination of diffusion probabilistic, flow-based, and generative adversarial network-based). Id.
10:54-57.
Again, Krisp makes the same arguments for why the ’496 Patent is directed at an abstract
idea. They fail for similar reasons as before: stripping away background noise while retaining
features key to what we recognize as a person’s voice (e.g., pitch, intonation, melody, stress) is not
a longstanding human process with an analogue equivalent; the patent is directed at particularized
design choices, methods, and ordered steps for achieving voice enhancement rather than at voice
Case 3:25-cv-05666-RS Document 72 Filed 02/23/26 Page 11 of 14

ORDER DENYING JUDGEMENT ON THE PLEADINGS
CASE NO. 25-cv-05666-RS
12
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
United States District Court
Northern District of California
enhancement itself; and the patent’s audio transformation is distinguishable from digital format
conversions that have been held patent ineligible.
D. ’457 and ’561 Patents
Krisp asserts claim 1 of the ’457 Patent is representative of all the asserted claims in the
’457 Patent, Real Time Correction of Accent in Speech Audio Signals, and ’561 Patent, also
named Real Time Correction of Accent in Speech Audio Signals, and that the other claims fail to
save either patent.7 Dkt. 40 at 21. Krisp argues both patents are directed to the abstract idea of
gathering, analyzing, and displaying certain results. Again, this is an oversimplification.
The ’457 and ’561 Patents describe correction of “chunks” of speech audio through the
extraction of certain acoustic and linguistic features and generation of new, corrected output. Dkt.
35-4 at 16:17-33 (claim 1). Specifically, the linguistic features identified and extracted include
those reflecting a standardized phonetic dictionary and those arising from a data-driven phonetic
library; the acoustic features include pitch, speed, and binary indicator of human speech; and the
features are identified by looking at both the chunk being corrected and a specific set of preceding
chunks. Id. at 16:51-55 (claim 5, dependent on claim 1); id. at 16:62-67 (claim 7, dependent on
claim 1); id. 16:25-26 (claim 1, “context associated with the input chunk”). The patent also
describes use of a neural network trained with an accent-reduction loss function and accent-
discriminator module to remove accent markers while preserving phonetic content. Id. at 16:55-16
(claim 6, dependent on claim 1), 18:46-52 (dependent claim 19). Moreover, the claims delineate
specific steps in which the audio should be processed, including when loudness should be
adjusted, id. at 17:58-65 (claim 13, dependent on claim 1), or ordered steps for generating
melspectograms, such as encoding linguistic features to generate hidden features, followed by
combining hidden features, acoustic features, and speaker embeddings, and lastly decoding, id. at
17:19-24 (dependent claim 10).
7 If the ’457 Patent is found patent eligible, Krisp’s motion also fails as to the ’561 Patent for
which it makes no particularized argument.
Case 3:25-cv-05666-RS Document 72 Filed 02/23/26 Page 12 of 14

ORDER DENYING JUDGEMENT ON THE PLEADINGS
CASE NO. 25-cv-05666-RS
13
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
United States District Court
Northern District of California
Krisp relies on Electric Power Group to argue these claims reflect an abstract idea without
an inventive concept to save it at Alice step two. Dkt. 40 at 23 (citing Elec. Power Grp., LLC v.
Alstom S.A., 830 F.3d 1350 (Fed. Cir. 2016)). However, the asserted claims at issue in Electric
Power Group differ dramatically from those at issue here. In Electric Power Group, the
representative claim details a method of detecting, analyzing, and displaying events and related
metrics on an electric power grid in real time. Elec. Power Grp., 830 F.3d at 1352. The metrics
and analyses to be displayed all reflect standard, non-inventive power system metrics except for a
“composite indicator of reliability,” which is not defined in any greater detail8 and for whose
derivation no process is described except for the “combination of one or more real time
measurements or computations of measurements[.]” Id. Further, no particular (let alone inventive)
method for real-time display is described. Id. at 1354 (“The advance they purport to make is a
process of gathering and analyzing information of a specified content, then displaying the results,
and not any particular assertedly inventive technology for performing those functions… limiting
the claims to the particular technological environment of power-grid monitoring is, without more,
insufficient to transform them into patent-eligible applications of the abstract idea at their core.”).
The unpublished opinion in Ubisoft Enterprises, on which Krisp also relies, is similarly
distinguishable. The patent at issue was for a guitar teaching program that displays finger
notations for a song, assesses whether the student has played the right notes, adjusts the difficulty
of the notations based on the student’s performance, and suggest songs or “mini-games” (portions
of songs or notations) based on the student’s skill level. Ubisoft Ent., S.A. v. Yousician Oy, 814 F.
App'x 588, 590 (Fed. Cir. 2020). Notably, “[n]othing in the claims or the specification of the ’852
patent disclose[d] a technological improvement over conventional methods.” Id. at 592.
The ’457 Patent, by contrast, generates an audio output whose features have been
specifically and inventively isolated and curated—which goes far beyond the display of standard
power metrics or musical notations and the admittedly abstract concept of gathering, analyzing,
and displaying results. Even if the ’457 Patent did not survive Alice step one, it survives Alice step
8 The description “an indicator of power grid vulnerability” provides no further detail. Id.
Case 3:25-cv-05666-RS Document 72 Filed 02/23/26 Page 13 of 14

ORDER DENYING JUDGEMENT ON THE PLEADINGS
CASE NO. 25-cv-05666-RS
14
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
United States District Court
Northern District of California
two because it describes “a combination of elements” that go to its specific method of accent
conversion such that “the patent in practice amounts to significantly more than a patent upon the
ineligible concept itself.” Alice, 573 U.S. 217–18 (internal citations and quotations omitted).
V. CONCLUSION
For the foregoing reasons, Krisp’s motion for judgment on the pleadings is denied as to all
counts on which Krisp moved.
IT IS SO ORDERED.
Dated: February 23, 2026
______________________________________
RICHARD SEEBORG
Chief United States District Judge
Case 3:25-cv-05666-RS Document 72 Filed 02/23/26 Page 14 of 14