Zentian v. Apple: How Claim Wording Broadened a Speech-Recognition Patent

A single pair of words can change the reach of a patent claim. In Zentian Ltd. v. Apple Inc., the Federal Circuit held that a claimed “feature vector” was not limited to quantities directly extracted from an audio stream. Because the claim expressly covered quantities that were “extracted and/or derived,” it also reached codewords produced through vector quantization.

That construction was decisive. Once the Patent Trial and Appeal Board treated the prior art’s codewords as feature vectors, it found every challenged claim of U.S. Patent No. 10,062,377 obvious. The Federal Circuit affirmed on September 25, 2026.

The decision is nonprecedential, but it is a useful drafting and PTAB lesson. Patent applicants often use broad connectors such as “and/or,” “derived,” “based on,” or “representative of” to avoid an unduly narrow claim. Those words can preserve valuable scope. They can also pull prior-art implementations into the claim and make an invalidity challenge easier.

What the speech-recognition patent claimed

The ’377 patent concerns speech-recognition systems. Its representative claim divides the work among programmable devices. One device calculates a feature vector from a digital audio stream during a defined time frame. A second device calculates distances that indicate similarity between the feature vector and acoustic-model states. A third device uses those distances, together with Hidden Markov Models and/or neural networks, to identify spoken words. The claim also recites a search stage that uses the calculated distances to identify words in a lexical tree.

The disputed language appeared in the claim’s definition of the feature vector. The vector could comprise multiple “extracted and/or derived” quantities from the digital audio stream.

Apple petitioned for inter partes review of claims 1–6. It relied on U.S. Patent No. 6,374,219 to Jiang, which describes converting feature vectors into codewords using vector quantization and a codebook. Vector quantization groups similar vectors to reduce computational demands. A codeword can serve as an index to a stored representative value for a cluster of data.

Zentian argued that Jiang’s codewords were not the original feature vectors. On that view, a distance calculated from a codeword was not a distance calculated from a feature vector, as the challenged claim required. Apple responded that the codewords were representative feature vectors and therefore fit the claim’s language.

The PTAB identified an implicit claim-construction dispute

Neither party initially offered an express construction of “feature vector,” but their invalidity positions depended on different meanings of that term. The PTAB therefore treated the dispute as one of claim construction.

Zentian’s proposed understanding was limited to what the Board described as “pure” feature vectors: quantities taken directly from the audio stream. Apple’s position also covered representations derived from the stream. The Board adopted the broader reading. It reasoned that the claim used both “extracted” and “derived,” and those different words should not be treated as though they meant the same thing.

Under that construction, a vector-quantized codeword could contain quantities derived from the audio stream even if it did not reproduce the originally extracted values. The Board found that Jiang disclosed the disputed limitation and ultimately held all six claims unpatentable for obviousness.

This procedural point matters well beyond speech recognition. Parties do not avoid claim construction merely by declining to propose formal definitions. When a petition and a patent-owner response map the same prior-art feature differently because they assign different scope to a claim term, the PTAB may need to resolve the embedded construction question. A party preparing an inter partes review or other PTAB proceeding should identify those disputes early and support its preferred reading with the full intrinsic record.

Why the Federal Circuit agreed that codewords were feature vectors

The Federal Circuit began with the claim itself. Because claim 1 said a feature vector could comprise quantities “extracted and/or derived” from the audio stream, the court presumed that “derived” added something beyond “extracted.” Limiting the term to values taken directly from the stream would make the additional word difficult to explain.

The court therefore agreed that a feature vector could include a representation derived from an audio stream. That conclusion covered Jiang’s codewords. Although a codeword might be identified by a single index, the evidence showed that it represented multiple quantities accessible through that index.

Zentian offered an analogy to René Magritte’s painting The Treachery of Images: a representation of a pipe is not itself a pipe, so a representation of a feature vector should not itself be a feature vector. The court found the analogy unpersuasive. Under the adopted construction and the Board’s factual findings, the codeword was both representative and a feature vector. In practical terms, it was closer to selecting one representative member for a group of similar data points.

Zentian also argued that codewords were determined in advance, before the relevant audio stream existed. The Board had found, however, that Jiang selected the codeword most closely resembling the computed vector for the actual stream. That selection produced a representative, quantized version derived from the stream. The Federal Circuit found substantial evidence supporting the Board.

The official Federal Circuit opinion in Zentian v. Apple affirmed the PTAB’s final written decision. The court did not create a categorical rule that every representation of data is the data itself. It interpreted this claim, whose own language expressly reached derived quantities, and reviewed the Board’s application of that construction to this record.

Broad claim language can protect more—and encounter more prior art

Patent drafting involves a recurring tension. Narrow language may distinguish the invention from known systems but leave competitors room to design around the claim. Broad language may cover more commercial variants but also sweep in more prior art.

Zentian shows that the tension can reside in ordinary-looking words. “Derived” can cover transformations, encodings, approximations, calculated values, selected representatives, or other downstream data, depending on the claim and specification. “And/or” can preserve alternatives, but it can also allow a reference that satisfies only one branch to meet the limitation.

The lesson is not to avoid those terms. The lesson is to use them deliberately. During patent application drafting, counsel should test each material term against at least three questions: what product variations should the claim cover, what earlier technologies might the wording capture, and what narrower dependent claims would preserve a commercially meaningful fallback?

The specification should then support the intended breadth. If “derived quantities” are meant to include certain transformations but exclude others, the application can describe the relevant relationships, processing steps, or properties. A definition that is too rigid may surrender needed flexibility, but silence can leave the boundary to be drawn later by an examiner, the PTAB, or a court.

The decision also illustrates a claim-construction contrast

The Federal Circuit does not automatically choose the broader construction in every Patent Office dispute. Context controls. In In re Incept, decided eight days before Zentian, the court held that the USPTO had read a “flow barrier” limitation too broadly during examination. The claim language and specification required a relationship that the Board’s construction had effectively erased.

The two results are consistent. In Incept, the intrinsic record supplied meaningful limits. In Zentian, the claim itself supplied an express breadth signal by including both extracted and derived quantities. Tucker Law’s overview of patent claim construction explains why the claim language, specification, and prosecution history must be evaluated together rather than one phrase at a time.

For patent owners, this comparison is a reminder to avoid arguing only that a construction feels too broad. The stronger argument identifies the particular words, grammatical relationships, definitions, embodiments, or prosecution statements that establish the proposed boundary. For challengers, the reverse is true: breadth should be anchored in the actual claim language and disclosure, not merely in the possibility that a term can bear a wider dictionary meaning.

Practical lessons for patent applicants and owners

A patent application should be reviewed not only for whether each claim reads on the preferred embodiment, but also for what else it might read on. That requires adversarial testing before filing. For data-processing inventions, the review should examine whether terms such as raw, extracted, derived, encoded, compressed, representative, transformed, and calculated are distinct, overlapping, or undefined.

Dependent claims can help manage that risk. A broader independent claim might cover extracted or derived data, while dependent claims can specify direct extraction, a particular transformation, a required data structure, a timing relationship, or a defined computational result. The dependent claims should reflect real commercial and technical alternatives—not limitations added solely for word count.

During prosecution, applicants should evaluate whether an examiner’s mapping exposes unintended claim breadth. Sometimes the best response is a construction argument grounded in the specification. Sometimes an amendment is more reliable. Arguments made to distinguish the reference should be considered for their later effect on infringement scope and enforceability.

Patent owners facing an IPR should also separate legal and factual issues. Claim construction is generally reviewed without deference, while findings about what a prior-art reference discloses receive substantial-evidence review. If the Board’s ruling could be characterized either way, the appellate strategy should address both. The Zentian court expressly noted that the distinction did not change the result because the Board prevailed under either standard.

What this means for technology companies

For software, artificial-intelligence, audio-processing, and data-compression patents, representations are often the technology. Systems routinely replace raw information with embeddings, tokens, hashes, indexes, quantized values, feature maps, or other derived forms. A claim that casually treats these categories as interchangeable may gain reach, but it may also collide with older systems that used a different label for the same functional relationship.

Portfolio reviews should therefore examine whether claim terminology tracks how the business and its competitors actually implement the technology. Tucker Law’s broader patent practice addresses prosecution, portfolio strategy, licensing, and enforcement as connected questions. The wording that helps obtain a patent will later shape invalidity, infringement, and deal diligence.

Likewise, companies assessing a competitor’s patent should not stop at the patent’s preferred embodiment. A properly construed claim may cover transformed or representative data even when the accused system does not use the original values. Conversely, the specification or prosecution record may impose a relationship that the claim’s shorthand does not reveal at first glance. Those issues often become central in patent litigation.

The bottom line

Zentian v. Apple is a concise illustration of how claim wording drives both scope and validity. The phrase “extracted and/or derived” helped make the claim broad enough to encompass codewords representing quantized data. That same breadth allowed the prior art to satisfy the limitation and supported an obviousness ruling against every challenged claim.

Patent applicants should choose breadth intentionally, build narrower fallback positions, and test important terms against the prior art before filing. Patent owners and challengers should surface implicit construction disputes early and develop both the intrinsic and technical evidence needed to resolve them.

If a claim-construction issue is affecting prosecution, an IPR, licensing diligence, or an infringement dispute, contact Tucker Law to evaluate the claim language, specification, prosecution history, and prior art together.

This article provides general information and is not legal advice. Patent scope and validity depend on the complete intrinsic record, prior art, procedural posture, and governing authority.