Federal Circuit Upholds PTAB Obviousness Finding in Apple IPR of Speech Recognition Patent

The U.S. Court of Appeals for the Federal Circuit (CAFC) issued a decision on September 25 in Zentian Ltd. v. Apple Inc., affirming a Patent Trial and Appeal Board (PTAB) final written decision that found all claims of a speech recognition patent unpatentable as obvious. The court rejected Zentian’s challenge to the PTAB’s construction of the term “feature vector” and affirmed the determination that codewords in a prior art reference fall within that term.

Apple Inc. filed a petition for inter partes review (IPR) of all six claims of U.S. Patent No. 10,062,377, owned by Zentian. The patent relates to speech recognition systems, and it describes converting the audio signal for each time frame of an audio input into a feature vector, then calculating a distance between that vector and states of an acoustic model to determine the probability that the vector corresponds to a given state.

Claim 1, which is representative of the issues on appeal, recites a speech recognition system with three programmable devices that calculate a feature vector, calculate distances to acoustic model states, and identify spoken words using those distances. Element 1[b] provides that the feature vector comprises a plurality of “extracted and/or derived quantities” from the digital audio stream, while element 1[g] requires that word identification use distances calculated “from a first feature vector.”

Apple’s petition relied on the U.S. Patent No. 6,374,219 (Jiang), asserting that the reference disclosed feature vectors and a further step of converting them into codewords through vector quantization and a codebook built from training data. Apple described a codeword as a representative feature vector and argued that Jiang disclosed calculating distances between codewords and acoustic model states.

Zentian’s patent owner response disputed that Jiang disclosed element 1[g], arguing that codewords are not the original underlying feature vectors. The Board noted that Zentian and Apple had not offered an explicit construction of “feature vector,” though their positions raised implicit ones, and it construed the term in light of that dispute. The Board framed the question as whether claimed feature vectors are limited to “pure” feature vectors, meaning quantities extracted from the audio stream as Zentian contended, or also include derived quantities that are representations of the stream as Apple contended. The Board adopted Apple’s position, relying on the reference in element 1[b] to extracted and/or derived quantities, found that Jiang disclosed element 1[g], and found all claims unpatentable as obvious.

Moreover, Zentian appealed, and the CAFC exercised jurisdiction under 28 U.S.C. Section 1295(a)(4)(A). The CAFC reviews claim construction de novo and any underlying fact findings for substantial evidence. Zentian argued that the Board erred by allowing “feature vector” to encompass codewords, which Zentian characterized as mere representations of the audio stream’s quantities, instead of limiting the term to quantities actually drawn from the stream.

The CAFC agreed with the Board, concluding that “the language of claim 1 itself resolves this dispute in Apple’s favor.” Since different claim terms are presumed to have different meanings, the CAFC presumed that “derived” in element 1[b] means something different from “extracted.” The CAFC therefore found no error in the Board’s conclusion that the term covers derived quantities that are representations of the audio stream, and affirmed that Jiang’s codewords fall within the term’s scope even if they do not comprise quantities actually extracted from the stream.

Zentian pointed to René Magritte’s painting The Treachery of Images, which depicts a tobacco pipe with text beneath it stating that it is not a pipe, to argue that representations of feature vectors are not feature vectors. The CAFC found a flaw in the analogy, since the Board found, with substantial evidence, that codewords represent the audio stream and are feature vectors under the construction. A codeword’s representative status “does not mean it is not also a ‘feature vector,’” according to the CAFC, which described a more fitting analogy, drawn from Apple’s brief, as choosing one pipe to represent a group of similar pipes.

Furthermore, Zentian also argued that codewords are determined before an audio stream is created and therefore cannot comprise quantities derived from it. The Board found that Jiang nonetheless uses predetermined codewords to produce vector quantized representations of the computed feature vectors. According to the CAFC, the selection of the codeword that most closely matches a given audio stream means the codeword is composed of quantities derived from that stream. The CAFC added that “Zentian has no persuasive response to this point.”

Zentian further argued that the Board’s construction removed the requirement that quantities be derived from the audio stream. The CAFC was unpersuaded, since the Board’s decision as a whole shows an understanding that the quantities must be derived from the stream. Zentian also argued that a codeword has only a single value and cannot comprise a plurality of quantities. Apple explained that a codeword may have a single-value index in a codebook while still comprising a plurality of quantities, and the Board shared that understanding.

Ultimately, the CAFC considered Zentian’s remaining arguments and found them unpersuasive. The court affirmed the Board’s final written decision determining claims 1 through 6 of the ‘377 patent unpatentable as obvious.

Image Source: Deposit Photos
Image ID:103763568
Copyright:bigfatnapoleon