This Is Not a Feature Vector: Zentian v. Apple

by Dennis Crouch

The Federal Circuit has affirmed the PTAB's cancellation of all six claims of Zentian's speech-recognition patent in a short nonprecedential opinion. Zentian Ltd. v. Apple Inc., No. 2024-2205 (Fed. Cir. Sept. 25, 2026).

Vector quantization: Jiang computes a feature vector for each short frame of audio, then swaps it for the closest entry in a codebook built in advance from training data. The codeword is itself a multi-component vector, a stand-in for a whole cluster of similar feature vectors. The swap gives up a little precision to save a lot of computation.

U.S. Patent No. 10,062,377 covers a speech-recognition system that splits the work across separate devices. The first device slices incoming audio into short time frames and converts each frame into a "feature vector," a set of numbers that places that sound as a point in N-dimensional space. A second device calculates the "distance" between each feature vector and the states of an acoustic model, and a third uses those distances (with Hidden Markov Models or neural networks and a lexical tree of words) to identify what was said.  The whole appeal turned on one term: "feature vector." Claim 1 requires requires word identification using distances "calculated from a first feature vector."

Apple's lead reference (Jiang) calculates those distances from vector-quantized "codewords" instead. The patentee Zentian argued that a codeword is not a feature vector. The Board and the court disagreed, relying mostly on the claim's own definition of a feature vector as comprising "extracted and/or derived quantities from said digital audio stream."

The case itself is largely a mill run, but it includes an interesting invocation of René Magritte's 1929 painting The Treachery of Images, the beautifully rendered tobacco pipe over the caption "Ceci n'est pas une pipe." In Zentian's words: "Just as a painting of a pipe is not a pipe, 'representations' of feature vectors are not feature vectors." Apple answered with a pipe analogy of its own -- that the prior art codeword was like "choosing one pipe to represent a group of similar pipes." Their point -- that the court accepted -- was that a codeword isn't a painting of a feature vector. Rather, It's a real feature vector picked from the codebook to stand in for a cluster of similar ones.

---

To continue reading, become a Patently-O member. Already a member? Simply log in to access the full post.