Preemptive Response Generation for Voice Search Latency

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Intelligent personal assistant systems experience significant response latencies due to inherent limitations in automatic speech recognition, IPA processing, and network delays, leading to undefined states for users and hindered workflow until complete processing is finished.

Innovation Solution

The system generates preemptive partial classification word candidates from the user's spoken request, applies a predictive component to extend these candidates, and performs a preliminary search to prepare preemptive responses, which are updated iteratively until the request is complete, allowing for immediate response once the request is finished.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If the system waits for complete speech recognition and IPA processing before responding, then response accuracy is improved, but response latency increases significantly

Engineering Contradiction:
Improveresponse accuracyVSAvoidresponse latency
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The system performs preliminary speech recognition processing on incoming utterances in real-time as they are spoken, generating partial recognition results before the user finishes speaking. This allows the system to prepare initial responses ahead of time, reducing the latency between speech completion and system response while maintaining accuracy through subsequent refinement of recognition results.

Inventive Principle:
Principle #10Preliminary action

2Speed

If the system processes speech input in real-time to reduce latency, then response speed is improved, but recognition accuracy may deteriorate

Engineering Contradiction:
Improveresponse speedVSAvoidrecognition accuracy
Core Design Contradiction:
SpeedVSMeasurement precision

Solution Approach 1:

The system maintains continuous speech recognition processing throughout the user's utterance, continuously updating partial recognition results as more speech data becomes available. This continuous processing ensures that the system always has the most current recognition results ready, improving response speed without sacrificing accuracy since the recognition process never stops refining its results.

Inventive Principle:
Principle #20Continuity of useful action

Solution Approach 2:

The system uses feedback from ongoing speech recognition to continuously update and refine partial recognition results. As the user continues speaking, the system compares incoming speech data with previously generated hypotheses and adjusts recognition results accordingly, ensuring that even real-time partial results maintain high accuracy while enabling fast response preparation.

Inventive Principle:
Principle #23Feedback

3Loss of time

If the system prepares preemptive responses during speech input, then response latency is reduced, but system complexity increases

Engineering Contradiction:
Improveresponse latencyVSAvoidsystem complexity
Core Design Contradiction:
Loss of timeVSDevice complexity

Solution Approach 1:

The system segments the speech recognition and response generation process into distinct stages: initial partial recognition during speech input, preliminary response preparation, and final response refinement. By dividing the overall process into these manageable segments, the system can work on multiple stages simultaneously without creating unmanageable complexity, allowing preemptive response preparation while maintaining system organization and controllability.

Inventive Principle:
Principle #1Segmentation

Data Source

PatentEP3092639B1A methodology for enhanced voice search experience
Publication Date: 2021.05.05 NUANCE COMMUNICATIONS INC
  • EP3092639B1 patent drawingFigure 1A
  • EP3092639B1 patent drawingFigure 1B
  • EP3092639B1 patent drawingFigure 1C

AI summary

Arrangements are described for reducing response latency in intelligent personal assistant applications. While receiving a user request, preemptive responses are automatically prepared for a received portion of the user request. Partial classification word candidates are generated for words in the received portion of the user request, and then a predictive component is applied to generate extended classification word candidates that include the partial classification word candidates and additional classification word candidates. A preliminary search is performed of the extended classification word candidates to prepare the preemptive responses. While the input request continues, the preemptive responses are updated, and when the input request ends, the prepared preemptive responses are used to respond to the user request.