N.D. California: Accent Translation Patents Survive Section 101 Patent Eligibility Attack

In Sanas.AI Inc. v. Krisp Technologies, Inc., Chief Judge Richard Seeborg of the U.S. District Court for the Northern District of California denied Krisp's Rule 12(c) motion challenging five Sanas patents on accent translation and speech processing. The court found the claims recite concrete technical methods – not abstract ideas – and distinguished the "format conversion" and "gather-analyze-display" cases Krisp invoked.

The court concluded that the asserted claims recite concrete technical methods rather than abstract ideas, distinguishing "format conversion" and "gather-analyze-display" cases. The order issued on February 23, 2026, in Case No. 3:25‑cv‑05666‑RS, permits Sanas' patent claims to proceed.

Procedural Posture and Scope of the Motion

The case pits two competitors in real-time speech technology. Sanas alleges Krisp infringes patents covering accent translation, accent mimicking, neural network-based voice enhancement and real-time accent correction. Krisp moved for judgment on the pleadings under Section 101, targeting five patents: the '550, '496 and '756 (supposedly just "translation") and '457 and '561 (supposedly "gather-analyze-display"). Sanas also asserts a '745 patent, but Krisp seemingly left that one alone. The court applied Alice on the papers.

Krisp's argument: Strip the claims to generic verbs (training, applying, synthesizing, converting) and call it all abstract.

The '550 patent? "Translation." The '756 patent on accent mimicking? A "longstanding human activity." The '496 patent's neural network voice enhancement? Same story.

The '457 and '561 patents? "Gather-analyze-display" per Electric Power Group.

Alice Step One: Specific Technical Methods, Not Abstract Results

The court did not buy it. Each patent, the court held, is directed to particular methods that improve speech processing technology – not merely claiming an abstract result. For the '550 patent: Claims align frames of input speech to phonemes, identify non‑text linguistic representations characteristic of an accent, map those to a second accent and perform frame‑wise mapping to generate output. Those aren't simply buzzwords. The court determined that they are concrete processing requirements enabling near real‑time conversion while preserving vocal characteristics.

The "format conversion" analogy failed. Those cases involved converting data between digital formats for device constraints. Here, the differences between input and output reflect accent features intelligible to humans – not device rendering. Different problem, different technology.

For the '756 patent (accent mimicking): The claims extract features, separate accent-associated traits from speaker identity and synthesize output preserving the speaker's voice while applying a different accent. The specification recites concrete features – phonetic patterns, prosody, articulation, intonation, timbre – and the claims order the processing. Though humans can imitate accents, that does not seem to matter when the claims delineate how to do it in a particular engineered way.

The '496 patent: a two‑network architecture operating on frame‑segmented audio. First neural network converts speech frames to low‑dimensional representations; second neural network generates target speech frames. Dependent claims specify diffusion-probabilistic, flow-based and GAN-based models. The court found this to be a particularized design – not an abstract aspiration to "strip noise" from audio.

For the '457 patent: specific acoustic and linguistic feature extraction, neural networks trained with an accent‑reduction loss function, ordered steps for generating mel spectrograms. The court added that even if step one failed, this combination would amount to an inventive concept at step two.

The argument for the '561 patent failed for the same reason: Krisp offered nothing beyond the same unsuccessful "gather-analyze-display" framing discussed above.

Two moves did the work. First: The court looked at how the claimed result is achieved. Claims that delineate steps showing how the technology achieves an improvement – or disclose improvements to the models – are more likely eligible. Claims that simply apply generic machine learning to a new data environment are not.

Second: The court refused to let Krisp strip away concrete claim elements through high-level abstraction. Reference to frames, phonemes and non-text linguistic units anchored the claims the way we saw in Enfish.

Electric Power Group didn't help Krisp. That case dealt with unadorned collection and display of known metrics – no described process for deriving the novel metric, no particular method for real‑time display. The claims here describe defined feature extraction, training objectives and ordered spectrogram generation. Different animal.

Practical Takeaways for AI and Speech‑Processing Claims

For AI and signal‑processing claims, the concrete "how" matters – as is a constant refrain. Here: specified representations (frames, phonemes, mel spectrograms), explicit alignment or mapping steps, defined feature taxonomies separating accent from speaker identity, model architectures and training regimes that change how the system operates. Where those elements are present and ordered to effect an improvement, courts are more likely to treat claims as eligible – not abstract results.

The order issued February 23, 2026 (N.D. California, Case No. 3:25-cv-05666-RS).

And Now, the Section 101 Blog's Oscar Preview That Nobody Asked For

The Academy Awards air Sunday. Ten Best Picture nominees and – unlike most Section 101 opinions – your family might actually want to discuss this with you.

This is a two-horse race, according to the pundits.

"One Battle After Another" is the film that will – and should – win Best Picture. Anderson adapted Thomas Pynchon, delivered a propulsive chase movie and, beneath all the sex, explosions and Sean Penn … stuff, made a movie about fatherhood. It is an aggressive and distinctly American film that somehow manages to be personal. The Academy should reward it.

"Sinners" (Ryan Coogler), with somehow 16 nominations is the darling, a blues/vampire epic set in 1930s Mississippi. But 16 nominations doesn't make a good movie. And, frankly, Sinners stunk once the vampires appeared. More Smoke and Stack, fewer vampires. The movie also didn't know how to end. It just … kept going. I love Buddy Guy in Chicago as much as the next Chicagoan, but what were we doing there?

I appear to be in the minority on Sinners. I am comfortable with that.

Here's the rest of the field, in my order (not what I think the academy will choose):

• "F1." Joseph Kosinski brings back the formula that made "Top Gun: Maverick" a smash: an aging star returns to high-speed competition, gorgeous cinematography, real stunts at 180 mph. Honestly, it's a bit like "Cars 3."

>Here's the thing: I did not like Maverick but very much enjoyed F1. F1 is like Maverick, but good. Why? I don't know. Maybe I just like Pitt more than Cruise. Or maybe I needed a movie that was less jingoistic. Not completely lacking jingoism – but just a bit less.

• "Marty Supreme." Josh Safdie directs Timothée Chalamet as an obnoxious, obsessively gifted ping-pong hustler in 1950s New York. Very entertaining. Underneath the chaos, it's a movie about a man to whom nothing matters more than himself – until he sees his son for the first time and everything changes. Nine nominations.
• "Bugonia." Yorgos Lanthimos and Emma Stone reunite for a conspiracy comedy about an amateur beekeeper who kidnaps a CEO he believes is an alien overlord. Very weird, very good, but it overstays its welcome by about 20 minutes. (Also, I didn't need that ending.)
• "Sentimental Value." Joachim Trier's Norwegian family drama deserves to be here. Stellan Skarsgård plays an aging auteur trying to reconnect with his estranged daughters and it very much lands. Unlike some of the others, I loved this ending. Nine nominations.
• "Train Dreams." Joel Edgerton plays a logger and railroad man in early-20th-century America. Edgerton does not say much, but the film never becomes a bore. Four nominations.
• "Frankenstein." Guillermo del Toro's monster movie is beautiful – and I enjoyed it. But I would've liked to enjoy it for about 45 fewer minutes.
• "Hamnet." Buckley was great. But this movie … this is not my kind of movie. [Frank Costanza voice] Eight nominations.
• ***"The Secret Agent." I haven't yet seen it.

And my other picks (not necessarily the Academy's):

• Best Director. Paul Thomas Anderson, One Battle After Another (some might say a long-overdue first win; I'd disagree)
• Best Actor. Leonardo DiCaprio, One Battle After Another (have to think Michael B. Jordan wins this, right?)
• Best Actress. Jessie Buckley, Hamnet (when she started wailing, I almost started crying, too – but mostly because I realized I still had another hour left in the movie)
• Best Supporting Actress. Inga Ibsdotter Lilleaas, Sentimental Value
• Best Supporting Actor. Stellan Skarsgård, Sentimental Value (ignore Penn; give it to Skarsgard)
• Best Original Screenplay. It's going to be Sinners (Coogler) but give me either Marty Supreme or Sentimental Value over it
• Best Adapted Screenplay. One Battle After Another (Anderson)

Enjoy Sunday. Back to Section 101 next week.