Amazon.com Inc v. Kaifi LLC

1. Case Identification

• Case #: IPR2025-00625
• Patent #: 8,930,196
• Filed: February 25, 2025
• Petitioner(s): Amazon.com, Inc.
• Patent Owner(s): Kaifi LLC
• Challenged Claims: 1-20

2. Patent Overview

• Title: Method and Apparatus for Recognizing Speech
• Brief Description: The ’196 patent discloses a two-step continuous speech recognition method for noisy environments. The system continuously listens for a "preset call command" (i.e., a wake word) and, upon recognition, activates a more computationally intensive continuous speech recognition unit to process a subsequent "actual command."

3. Grounds for Unpatentability

Ground 1: Obviousness over Bou-Ghazale and Konopka - Claims 1-5, 8-9, 16-17, and 19-20 are obvious over Bou-Ghazale in view of Konopka.

• Prior Art Relied Upon: Bou-Ghazale (a 2000 IEEE article) and Konopka (Application # 2008/0059188).
• Core Argument for this Ground:
  • Prior Art Mapping: Petitioner argued that Bou-Ghazale taught the core invention: a two-step voice activation system that continuously listens for a pre-defined activation command and, once recognized, activates a main speech recognition system to process "further commands." This two-stage process was intended to reduce computational load on a device, such as a personal communication device. Bou-Ghazale’s system used a hidden Markov model (HMM) based component that performed a "trace-back" at every frame to find the result string with the highest likelihood score.
  • Motivation to Combine: Konopka taught that "token passing" was a "well-known approach" to improve HMM-based systems by providing a memory-efficient solution to the problem of storing back-pointers during the trace-back process. Petitioner asserted a Person of Ordinary Skill in The Art (POSITA) would combine Konopka’s well-known token passing technique with Bou-Ghazale’s HMM-based system to solve the known problem of high memory requirements associated with traditional back-tracing, thereby making Bou-Ghazale's system more efficient.
  • Expectation of Success: A POSITA would have a reasonable expectation of success because token passing was specifically intended to improve HMM-based speech recognition systems like Bou-Ghazale's, and the combination would not change the intended functionality of either system.

Ground 2: Obviousness over Bou-Ghazale, Konopka, and Lee - Claims 6-7 are obvious over Bou-Ghazale in view of Konopka and Lee.

• Prior Art Relied Upon: Bou-Ghazale, Konopka, and Lee (a 2005 article in Speech communication).
• Core Argument for this Ground:
  • Prior Art Mapping: This ground built upon the combination of Bou-Ghazale and Konopka to address limitations in dependent claims 6 and 7 related to computing log-likelihood ratios using phoneme and anti-phoneme models. Petitioner contended that Lee taught a technique for verifying voice codes in memory-limited systems using a phoneme model (representing the voice code) and an anti-phoneme model (representing sounds statistically far from the voice code) to reject non-speech noises and out-of-vocabulary words. Lee's technique of subtracting scores from these two models was argued to map directly to the claimed method of calculating "another log likelihood ratio."
  • Motivation to Combine: A POSITA would be motivated to incorporate Lee’s technique into the Bou-Ghazale/Konopka system to further the shared goals of accurately recognizing commands, rejecting noise, and reducing memory requirements. Lee’s method was advantageous because it did not require a separately trained "garbage model," thereby saving memory space, which directly aligned with Bou-Ghazale’s stated goal of implementation on low-memory devices.
  • Expectation of Success: Success was predictable because Lee’s technique was designed for HMM-based systems and would improve the combined system's noise rejection and memory efficiency without altering its fundamental two-step operation.

Ground 3: Obviousness over Bou-Ghazale, Konopka, and Takiguchi - Claims 10-13 and 18 are obvious over Bou-Ghazale in view of Konopka and Takiguchi.

• Prior Art Relied Upon: Bou-Ghazale, Konopka, and Takiguchi (Application # 2006/0122832).

• Core Argument for this Ground:

  • Prior Art Mapping: This ground addressed apparatus claims reciting an "adaptive filter." Petitioner argued that while Bou-Ghazale taught using a simple "band-pass filter" with a fixed frequency range, Takiguchi taught a superior "adaptive filter" for speech enhancement. Takiguchi’s filter updated its coefficients for each speech portion to more effectively reduce extemporaneous noise signals.
  • Motivation to Combine: A POSITA would have found it obvious to replace the basic band-pass filter in Bou-Ghazale’s system with the improved adaptive filter taught by Takiguchi. This modification represented a simple substitution of one known element for another to obtain predictable results—namely, improved speech recognition accuracy in the noisy environments that Bou-Ghazale’s system was designed to operate in.
  • Expectation of Success: There was a high expectation of success, as applying an improved filtering technique to a speech recognition system was a routine and predictable design choice for a POSITA seeking to enhance performance.

• Additional Grounds: Petitioner asserted an additional obviousness challenge (Ground 4) for claims 14-15 based on the combination of Bou-Ghazale, Konopka, Takiguchi, and Lee, which relied on the teachings and motivations established in the preceding grounds.

4. Arguments Regarding Discretionary Denial

• Petitioner argued that discretionary denial under Fintiv would be inappropriate. The co-pending district court litigation was asserted to be in its early stages with minimal investment, as key events like claim construction and expert discovery were not scheduled until late 2025. Petitioner noted that the trial date was speculative and that the petition presented compelling evidence of unpatentability, weighing against denial. Petitioner also argued that denial under §325(d) was unwarranted as none of the relied-upon prior art references were previously considered by the USPTO during prosecution.

5. Relief Requested

• Petitioner requests institution of IPR and cancellation of claims 1-20 of the ’196 patent as unpatentable.