Prior art

Analysis of Prior Art for U.S. Patent 11,929,073

Here is an analysis of the most relevant prior art cited during the examination of U.S. Patent 11,929,073. This analysis focuses on references that could potentially anticipate the independent claims of the patent under 35 U.S.C. § 102.

The core of the invention in patent 11,929,073 is a two-stage arbitration process for a hybrid speech recognition system. The first stage, or "short-circuit," decides whether to accept a locally-generated recognition result immediately or to wait for a result from a more powerful cloud-based system. This initial decision is made before the cloud result is even received.

---

Key Prior Art References

1. U.S. Patent Application Publication No. US 2013/0346078 A1

• Full Citation: US 2013/0346078 A1, "Mixed model speech recognition," assigned to Google Inc.
• Publication Date: December 26, 2013 (Filed June 26, 2012)
• Brief Description: This reference describes a hybrid speech recognition system that uses both a local, on-device recognizer and a server-based (cloud) recognizer. It details a process where both recognizers generate transcription hypotheses for a given utterance. The system then selects one of these transcriptions based on various signals, such as confidence scores. It explicitly mentions the trade-off between the low latency of the local recognizer and the potentially higher accuracy of the server-based recognizer.
• Potential Anticipation of Claim(s): 1, 14, 15
  • This reference appears to be highly relevant. It discloses a hybrid system that solicits results from both a local device and a cloud service, a core concept of the claims. The '078 application describes selecting a final transcription from the outputs of the two systems. While it may not explicitly use the term "short-circuit" or detail the exact two-stage decision logic of the '073 patent (deciding whether to wait based only on the local result first), the fundamental components and the overall goal of arbitrating between a local and cloud result are present. An argument for anticipation could be made that determining whether to use the faster local result or wait for the more accurate cloud result is an inherent trade-off described in this prior art, even if the specific implementation differs.

2. U.S. Patent No. 10,186,262 B2

• Full Citation: US 10,186,262 B2, "System with multiple simultaneous speech recognizers," assigned to Microsoft Technology Licensing, LLC.
• Publication Date: January 22, 2019 (Filed July 31, 2013)
• Brief Description: This patent details a system where multiple speech recognizers operate simultaneously on the same audio input. These recognizers can differ in their models, vocabularies, or contexts (e.g., one general-purpose, one domain-specific). The system then includes a "recognition arbiter" that selects the best result from the multiple outputs. The arbiter can use confidence scores and other metadata to make its selection.
• Potential Anticipation of Claim(s): 1, 14, 15
  • Like the '078 application, the '262 patent discloses the core idea of using multiple speech recognizers (which could be local and cloud-based) and an arbiter to select the best result. The concept of "simultaneous" recognition aligns with the '073 patent's process of soliciting both local and cloud results in parallel. The "recognition arbiter" in the '262 patent performs a similar function to the arbitrator in the '073 patent. The key question for anticipation would be whether the '262 patent teaches or suggests the specific "short-circuit" logic of making a preliminary decision to select the local result before the second (cloud) result is available, based solely on the quality of that first result.

3. U.S. Patent Application Publication No. US 2018/0342236 A1

• Full Citation: US 2018/0342236 A1, "Automatic multi-performance evaluation system for hybrid speech recognition," assigned to Mediazen, Inc.
• Publication Date: November 29, 2018 (Filed in Korea October 11, 2016)
• Brief Description: This document describes a system for evaluating and selecting between results from an embedded (local) speech recognition engine and a server-based (cloud) speech recognition engine. It explicitly discusses a hybrid system that sends speech data to both engines. The system then evaluates the results and determines a final recognition based on a "selection policy," which can factor in performance and confidence.
• Potential Anticipation of Claim(s): 1, 14, 15
  • This reference is also highly relevant as it directly addresses the arbitration between local and cloud speech recognition results. It teaches acquiring speech data, soliciting results from both a computing device and a cloud service, and then applying a policy to select the final result. The anticipation argument would hinge on whether its "selection policy" anticipates the specific two-step decision process claimed in the '073 patent, particularly the initial, time-saving decision to accept the local result without waiting for the cloud response.

4. U.S. Patent Application Publication No. US 2019/0043496 A1

• Full Citation: US 2019/0043496 A1, "Distributed speech processing," assigned to Intel Corporation.
• Publication Date: February 7, 2019 (Filed September 28, 2017)
• Brief Description: This application describes a distributed speech processing system where a client device can perform initial speech processing and a server can perform more complex processing. It discusses a "decider" module that can determine where speech processing should occur (client, server, or both) based on factors like network conditions, device capabilities, and the complexity of the query.
• Potential Anticipation of Claim(s): 1, 14, 15
  • The '496 application teaches a system that arbitrates between local and server-based processing. Its "decider" module performs a function analogous to the arbitrator in the '073 patent. It discloses the concept of making a determination based on the initial data from the client side. This could be interpreted as anticipating the "short-circuit" decision. For example, if the decider determines the query is simple and can be handled locally with high confidence, it might preemptively select the local result without waiting for a full server-side process, which aligns closely with the logic in Claim 1 of the '073 patent.