Dual-Mode Speech Recognition Model for Latency-Accuracy Tradeoffs

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing speech recognition technologies face challenges in achieving both low-latency streaming recognition and high-accuracy contextual recognition, with separate models required for each mode, leading to increased computational costs and complexity.

Innovation Solution

A dual-mode machine-learned speech recognition model that can switch between streaming and contextual recognition modes, utilizing a dual-mode layer architecture and a combined loss function to distill contextual accuracy into streaming mode, reducing the need for separate models.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If separate models are used for streaming and contextual recognition, then each mode can be optimized independently, but device complexity and computational costs increase

Engineering Contradiction:
Improverecognition accuracyVSAvoidmodel complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent combines streaming and contextual recognition capabilities into a single unified model. The model includes both a streaming recognition path that processes speech in real-time and a contextual recognition path that processes complete speech segments, allowing both functions to coexist in one model structure rather than requiring separate models for each mode.

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The unified model is designed to perform multiple functions: it can operate in streaming mode for low-latency applications and in contextual mode for high-accuracy applications. The model structure incorporates components that enable both recognition approaches, making it a universal solution that adapts to different operational requirements without needing separate specialized models.

Inventive Principle:
Principle #6Universality (Multi-functionality)

2Loss of time

If streaming recognition is used, then low-latency performance is achieved, but prediction accuracy deteriorates

Engineering Contradiction:
ImprovelatencyVSAvoidprediction accuracy
Core Design Contradiction:
Loss of timeVSMeasurement precision

Solution Approach 1:

The model segments the recognition process into two distinct paths: a streaming path that processes speech incrementally for low-latency applications, and a contextual path that processes complete speech segments for high-accuracy applications. This segmentation allows the system to choose the appropriate path based on latency requirements while maintaining both capabilities within the unified model.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The model dynamically switches between streaming and contextual recognition modes depending on the application requirements. The streaming mode is activated when low-latency is critical, while the contextual mode is activated when maximum accuracy is needed. This dynamic adaptability allows the single model to optimize performance for different operational scenarios.

Inventive Principle:
Principle #15Dynamics

3Measurement precision

If contextual recognition is used, then prediction accuracy is improved, but processing time increases

Engineering Contradiction:
Improveprediction accuracyVSAvoidprocessing time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The model segments the recognition process into two distinct paths: a streaming path that processes speech incrementally for low-latency applications, and a contextual path that processes complete speech segments for high-accuracy applications. This segmentation allows the system to choose the appropriate path based on latency requirements while maintaining both capabilities within the unified model.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The model dynamically switches between streaming and contextual recognition modes depending on the application requirements. The streaming mode is activated when low-latency is critical, while the contextual mode is activated when maximum accuracy is needed. This dynamic adaptability allows the single model to optimize performance for different operational scenarios.

Inventive Principle:
Principle #15Dynamics

4Reliability

If multiple separate models are maintained, then mode-specific optimization is achieved, but computational costs and resource utilization increase

Engineering Contradiction:
Improvemode-specific performanceVSAvoidcomputational cost
Core Design Contradiction:
ReliabilityVSUse of energy by moving object

Solution Approach 1:

The patent combines streaming and contextual recognition capabilities into a single unified model. The model includes both a streaming recognition path that processes speech in real-time and a contextual recognition path that processes complete speech segments, allowing both functions to coexist in one model structure rather than requiring separate models for each mode.

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The unified model is designed to perform multiple functions: it can operate in streaming mode for low-latency applications and in contextual mode for high-accuracy applications. The model structure incorporates components that enable both recognition approaches, making it a universal solution that adapts to different operational requirements without needing separate specialized models.

Inventive Principle:
Principle #6Universality (Multi-functionality)

Data Source

PatentEP4154248B1Systems and methods for training dual-mode machine-learned speech recognition models
Publication Date: 2026.03.11 GOOGLE LLC
  • EP4154248B1 patent drawingFigure 1A
  • EP4154248B1 patent drawingFigure 1B
  • EP4154248B1 patent drawingFigure 1C

AI summary

Systems and methods of the present disclosure are directed to a computing system, including one or more processors and a machine-learned multi-mode speech recognition model configured to operate in a streaming recognition mode or a contextual recognition mode. The computing system can perform operations including obtaining speech data and a ground truth label and processing the speech data using the contextual recognition mode to obtain contextual prediction data. The operations can include evaluating a difference between the contextual prediction data and the ground truth label and processing the speech data using the streaming recognition mode to obtain streaming prediction data. The operations can include evaluating a difference between the streaming prediction data and the ground truth label and the contextual and streaming prediction data. The operations can include adjusting parameters of the speech recognition model.