Google LLC v. Buffalo Patents LLC

1. Case Identification

• Case #: IPR2023-01386
• Patent #: 6,904,405
• Filed: September 12, 2023
• Petitioner(s): Google LLC
• Patent Owner(s): Buffalo Patents, LLC
• Challenged Claims: 1-12

2. Patent Overview

• Title: System and method for performing multiple different types of message recognition using a shared language model
• Brief Description: The ’405 patent discloses systems and methods for performing message recognition for multiple input types (e.g., speech and handwriting) using a unified message model that includes a shared language model and input-specific models. The system aims to improve accuracy by allowing the shared language model to be trained based on user corrections from either input type.

3. Grounds for Unpatentability

Ground 1: Claims 1-3, 6-9, and 12 are obvious over Waibel (incorporating Suhm, Vo, and Manke) in view of Roberts.

• Prior Art Relied Upon: Waibel (a 1997 IEEE publication), Suhm (a 1995 publication), Vo (a 1996 publication), Manke (a 1994 publication), and Roberts (Patent 5,027,406).
• Core Argument for this Ground:
  • Prior Art Mapping: Petitioner asserted that Waibel, which explicitly incorporates the teachings of Suhm, Vo, and Manke, discloses a multimodal system for generating text responsive to different message inputs (speech and handwriting). Waibel’s system was argued to include a speech recognizer (from Suhm) and a handwriting recognizer (from Manke) that feed into a joint multimodal interpreter (from Vo). Petitioner contended this combination discloses the claimed "unified message model" by teaching a first model specific to speech recognition (e.g., Suhm's acoustic models), a second model specific to handwriting recognition (e.g., Manke's HMM-based models), and a shared language model. The shared language model was argued to be present or obvious because Manke teaches using language models "originally defined for speech recognition" to achieve "synergy effects," and Vo describes using dialog frames to merge information from both input types, which functions as a dynamic, shared language model.
  • Motivation to Combine: Petitioner argued that while Waibel discloses the core system architecture, it does not explicitly teach that the shared language model is trainable based on user corrections. Roberts was introduced to supply this limitation. Roberts teaches a speech recognition system where user corrections of misrecognized words are used to update and train both acoustic and language models, thereby improving system performance over time. A POSITA would combine Roberts with Waibel to incorporate this known technique for improving accuracy into Waibel’s multimodal system, as both systems rely on similar underlying technologies like HMMs and GUI-based error correction.
  • Expectation of Success: A POSITA would have had a reasonable expectation of success because Roberts’s method of training HMM-based models is directly applicable to the HMM-based recognizers disclosed in Waibel (via Suhm and Manke). The combination was presented as the application of a known improvement technique to a similar existing system to achieve predictable results.

Ground 2: Claims 1-3 and 7-9 are obvious over Roth in view of Roberts.

• Prior Art Relied Upon: Roth (WO 92/05517) and Roberts (Patent 5,027,406).

• Core Argument for this Ground:

  • Prior Art Mapping: Petitioner argued that Roth discloses an "Audio-Augmented Handwriting Recognition" system that generates text responsive to both speech and handwriting inputs. Roth's system was asserted to meet the limitations of a "unified message model" by using a pronunciation-by-rule algorithm and a template store as a "shared language model" to correlate outputs from both recognizers. Roth further discloses speech-specific models (e.g., speech templates) and handwriting-specific models (disclosed as being conventional). The system includes a speech recognizer (microphone input) and a handwriting recognizer (touchpad input) that work together to produce a text output.
  • Motivation to Combine: As with the Waibel combination, Petitioner contended Roth does not explicitly teach that its shared language models are trainable based on user corrections. Roberts was again introduced to teach this feature. A POSITA would be motivated to use Roberts's training method to improve the accuracy of Roth's recognizers. Specifically, by using Roth's existing user correction mechanism to update its shared language models (e.g., the pronunciation rules or speech templates), the system would become more accurate and efficient, which are the exact benefits described in Roberts.
  • Expectation of Success: Success was expected because Roth’s user correction methods are compatible with the training methods taught by Roberts. Incorporating a trainable bigram language model, as taught by Roberts, into Roth's existing ranking methodology would be a straightforward modification to improve recognition of multi-word inputs, a predictable enhancement.

• Additional Grounds: Petitioner asserted additional obviousness challenges, including combinations adding Verrier (EP 0622724A2), which teaches a stylus with an integrated microphone, to the primary grounds to address limitations in dependent claims 4, 5, 10, and 11. Another ground added Vo to the Roth and Roberts combination to teach a semantic model for claims 6 and 12.

4. Arguments Regarding Discretionary Denial

• §314(a) (Fintiv): Petitioner argued against discretionary denial under Fintiv, stating that the parallel district court litigation against a third party (Motorola) was stayed, discovery had not commenced, and no trial date was set. Given the early stage of the co-pending litigation, Petitioner contended that an IPR would be a more efficient resolution.
• §325(d): Petitioner argued against denial under §325(d), asserting that the primary prior art references relied upon in the petition (Waibel, Roberts, Roth, Suhm, Vo, Manke) were never considered by the PTO during the original prosecution of the ’405 patent. Petitioner noted that while a different patent by Waibel was cited by the examiner, it described a different system and was not combined with Roberts, meaning the core arguments and prior art combinations presented in the petition are new to the PTO.

5. Relief Requested

• Petitioner requests institution of an inter partes review and cancellation of claims 1-12 of the ’405 patent as unpatentable under 35 U.S.C. §103.