Partial-Hypothesis Streaming Action Fulfillment Before Speech Endpoint

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing digital assistant interfaces on mobile devices experience user perceived latency in fulfilling user queries due to the need to wait for the endpoint of an utterance before initiating sub-actions, despite advancements in on-device speech recognition that reduce latency for transcription.

Innovation Solution

Implementing a system that generates intermediate speech recognition results during an utterance to perform sub-actions related to the final action before the user finishes speaking, leveraging partial hypotheses to initiate actions in real-time.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If the system waits for the endpoint of an utterance before initiating sub-actions, then the accuracy of speech recognition is improved, but the user perceived latency increases

Engineering Contradiction:
Improvespeech recognition accuracyVSAvoiduser perceived latency
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The system performs preliminary speech recognition processing during the utterance to generate intermediate results and partial hypotheses before the endpoint is reached. This allows sub-actions to be initiated in advance based on incomplete but progressively improving speech recognition data, thereby reducing user perceived latency while maintaining acceptable accuracy through continuous refinement.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The system performs speech recognition and action initiation based on partial speech data rather than waiting for complete utterances. By accepting intermediate speech recognition results with varying confidence levels, the system can start processing sub-actions earlier, trading off some precision for reduced latency, and then refine or correct actions as more complete speech data becomes available.

Inventive Principle:
Principle #16Partial or excessive action

2Loss of time

If the system performs speech recognition on-device, then the latency for transcription is reduced, but the reliability of action fulfillment is worsened due to slow or unreliable local applications

Engineering Contradiction:
Improvetranscription latencyVSAvoidaction fulfillment reliability
Core Design Contradiction:
Loss of timeVSReliability

Solution Approach 1:

The system introduces an intermediary layer of cloud-based services that receives intermediate speech recognition results and coordinates the execution of sub-actions across multiple applications. This intermediary cloud service acts as a reliable orchestrator that can manage complex action sequences involving multiple applications, compensating for the unreliability of individual local applications while preserving the low-latency benefits of on-device speech recognition.

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

The system segments the action fulfillment process into multiple independent sub-actions that can be executed in parallel or sequence. By breaking down the overall task into smaller, manageable sub-actions, the system can distribute execution across different local applications and cloud services, improving overall reliability through modular execution while maintaining the speed advantages of on-device processing.

Inventive Principle:
Principle #1Segmentation

Data Source

PatentEP4244849B1Streaming action fulfillment based on partial hypotheses
Publication Date: 2025.08.20 GOOGLE LLC
  • EP4244849B1 patent drawingFigure 1
  • EP4244849B1 patent drawingFigure 2A
  • EP4244849B1 patent drawingFigure 2B

AI summary

A method (300) for streaming action fulfillment receives audio data (14) corresponding to an utterance (20) where the utterance includes a query (22) to perform an action (24) that requires performance of a sequence of sub-actions (26) in order to fulfill the action. While receiving the audio data, but before receiving an end of speech condition (214), the method processes the audio data to generate intermediate automated speech recognition (ASR) results (212), performs partial query interpretation on the intermediate ASR results to determine whether the intermediate ASR results identify an application type needed to perform the action and, when the intermediate ASR results identify a particular application type, performs a first sub-action in the sequence of sub-actions by launching a first application (118) to execute on the user device where the first application is associated with the particular application type. The method fulfills performance of the action.