Sanas.AI v. Krisp Technologies — Court Finds AI Accent Conversion Patents Survive Section 101 Challenge

Background

Sanas.AI and Krisp Technologies are direct competitors in the market for real-time audio-enhancing technology. Sanas, founded by three former Stanford classmates, developed patented technology that converts accents in real-time during phone calls and video conferences, using a student-teacher machine learning architecture and parallel data generation techniques to translate speech fast enough for natural conversation. Krisp, originally focused on background noise suppression, entered the accent conversion market in 2023 after extensive partnership discussions with Sanas fell apart.

Sanas sued Krisp for infringement of six patents covering its accent conversion technology, along with trade secret and other claims. Sanas alleges that Krisp’s accent conversion patents actually cite Sanas’ earlier applications and build upon information shared by Sanas employees during the failed partnership discussions. Krisp counterclaimed for infringement of its own noise suppression patent.

After the court denied earlier motions to dismiss, Krisp moved for judgment on the pleadings under Rule 12(c), arguing that Sanas’ patents are directed to abstract ideas—either “translation” (converting audio from one form to another) or the generic process of gathering, analyzing, and displaying data—and are therefore ineligible under 35 U.S.C. § 101.

The Court’s Holding

Chief Judge Seeborg denied Krisp’s motion as to all five challenged patent groups, finding that each survives the Alice framework at Step 1.

The ’550 Patent (Real-time Accent Conversion Model): The court found this patent is directed at specific methods for converting accents—identifying non-text linguistic units of speech, mapping them between accents using trained machine learning models, and generating audio output—rather than at the abstract concept of “translation.” The court rejected Krisp’s analogy to digital format conversion cases, noting that accent conversion involves differences “intelligible and meaningful to humans apart from the devices rendering them,” unlike converting video between digital codecs.

The ’496 Patent (Neural Network-Based Voice Enhancement): Similarly, the court found this patent claims specific technical mechanisms—real-time accent conversion using particular neural network architectures—not generic machine learning. The court cited the Federal Circuit’s Recentive Analytics v. Fox Corp. (2025), which held that claims “delineating steps through which the machine learning technology achieves an improvement” are patent eligible, while claims disclosing only “the application of generic machine learning to new data environments” are not.

The ’457 and ’561 Patents (Real Time Accent Correction): Krisp argued these patents merely claim gathering, analyzing, and displaying data. The court disagreed, finding that the patents disclose specific correction mechanisms rather than generic data processing steps.

The ’756 Patent (Accent Mimicking): The court also rejected Krisp’s challenge to this patent, finding it claims specific systems and methods for real-time accent mimicking that go beyond abstract concepts.

Throughout the analysis, the court warned against “describing the claims at a high level of abstraction, divorced from the claim language itself,” which would constitute reversible error under Federal Circuit precedent.

Key Takeaways

• AI patents that claim specific ML architectures can survive § 101. The court drew a clear line from Recentive Analytics: patents that delineate specific steps through which ML achieves an improvement are eligible; patents that merely apply generic ML to new data are not. Sanas’ patents fell on the right side because they claimed particular techniques—frame alignment, phoneme mapping, student-teacher training—rather than AI in the abstract.
• Accent conversion is not “digital format conversion.” The court rejected the analogy to prior cases involving conversion between video codecs or digital formats, finding that converting human-intelligible speech characteristics between accents is fundamentally different from converting data between machine-readable formats.
• Competitors who enter a market after failed partnership talks face heightened risk. The backdrop of Sanas and Krisp’s failed collaboration—after which Krisp entered the accent conversion market with patents that cite Sanas’ applications—adds a trade secret dimension. The court previously denied Krisp’s motion to dismiss the non-patent claims as well.

Why It Matters

This ruling provides early guidance on how courts will evaluate patent eligibility for AI-powered speech and audio technology—a rapidly growing market driven by remote work, customer service automation, and communications tools. As companies race to patent AI innovations, the line between abstract ideas and patentable technical improvements remains the most contested question in patent law. Chief Judge Seeborg’s analysis suggests that courts will look favorably on AI patents that describe specific architectures and processing steps rather than claiming outcomes or invoking generic ML concepts.

For the broader AI industry, the case also highlights the risks of collaborative discussions that don’t result in a deal: Sanas alleges that Krisp’s patents are built on information shared during their partnership talks, a pattern increasingly common in the tech sector where potential partners become competitors.