Zentian v. Apple — Speech-Recognition Patent Claims Remain Obvious

Background

Zentian’s U.S. Patent No. 10,062,377 concerns speech-recognition systems. The system converts portions of a digital audio stream into “feature vectors,” compares those vectors with acoustic-model states, and uses the resulting distances to help identify spoken words. Apple challenged all six claims in inter partes review.

The Patent Trial and Appeal Board found the claims obvious over a prior-art reference known as Jiang. Jiang used vector quantization: it compared a computed vector with entries in a codebook and selected the closest stored “codeword.” The central dispute on appeal was whether those codewords could satisfy the patent’s “feature vector” limitation.

The Court’s Holding

The Federal Circuit affirmed the Board’s construction and obviousness determination. Claim 1 said a feature vector could comprise quantities “extracted and/or derived” from the audio stream. Giving “derived” meaning distinct from “extracted,” the panel held that the term was not limited to raw or “pure” quantities taken directly from the audio. It also encompassed representative quantities derived through processing, including Jiang’s codewords.

Zentian compared a codeword to René Magritte’s painting of a pipe: a representation of a thing is not the thing itself. The panel found the analogy inapt. Substantial evidence supported the Board’s finding that a selected codeword was both representative of the audio-derived vector and itself a feature vector under the claim language. A closer analogy, the court said, was choosing one pipe to represent a group of similar pipes.

The court also rejected the argument that a precomputed codeword could not be derived from a later audio stream. Selecting the closest codeword based on the incoming audio produced a representative vector derived from that stream. And although a codeword could be identified by a single index, the indexed entry still represented multiple quantities, satisfying the claim’s plurality requirement.

Key Takeaways

• Different words in a claim—here, “extracted” and “derived”—are ordinarily presumed to carry different meanings.
• A stored representation can fall within a claim term even when it is selected from a precomputed codebook.
• Courts test analogies against the patent’s actual language and the Board’s factual findings, not their rhetorical appeal.

Why It Matters

The decision illustrates how a few words in a claim can determine the outcome of an obviousness challenge. By claiming quantities that were “derived” from audio, Zentian used language broad enough to cover processed representations in the prior art. Patent drafters must balance breadth against the risk that broad terminology will sweep in known techniques.

This is a separate appeal from an earlier Apple–Zentian dispute involving a different speech-recognition patent. LexSummary covered that earlier ruling here. Together, the decisions show the Federal Circuit closely policing both overly narrow Board assumptions and overly narrow patent-owner claim constructions.