Modular Deep Learning Acoustic Model for Context Adaptation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current acoustic models face challenges in accurately processing audio signals across a variety of contexts, as a single model is less accurate for broad contexts and poorly performs when encountering signals outside its specific training context, due to data scarcity and difficulty in training separate models for each context.

Innovation Solution

A modular deep learning acoustic model is developed, comprising shared layers and context-specific sub-modules that can be swapped in and out without additional runtime search cost, allowing the model to adapt to specific acoustic conditions such as speaker variations, accents, and environmental factors.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Device complexity

If a single acoustic model is used to process audio signals across a variety of contexts, then the device complexity is reduced, but the measurement precision deteriorates due to inability to accurately identify sounds across different environments and speakers

Engineering Contradiction:
Improvemodel structureVSAvoidspeech recognition accuracy
Core Design Contradiction:
Device complexityVSMeasurement precision

Solution Approach 1:

The acoustic model is segmented into a shared backbone and multiple context-specific sub-modules. Each sub-module is trained on specific acoustic contexts (different speakers, environments, accents) while sharing the common feature extraction layers. This segmentation allows the system to maintain a relatively simple overall structure while achieving high precision for different contexts through specialized sub-modules.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The shared layers of the acoustic model provide universal feature extraction capabilities that work across all acoustic contexts. These universal layers are combined with context-specific sub-modules, allowing the same base model to serve multiple functions and adapt to different speech recognition scenarios without requiring completely separate models for each context.

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

2Measurement precision

If separate acoustic models are trained for each specific acoustic context, then the measurement precision improves for that context, but the device complexity increases due to multiple models needed for different contexts

Engineering Contradiction:
Improvespeech recognition accuracyVSAvoidmodel structure
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

Instead of maintaining completely separate models for each context, the system segments the model into shared and context-specific components. This reduces complexity by eliminating redundant feature extraction layers while preserving context-adaptive capabilities through specialized sub-modules for each acoustic environment and speaker type.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

Multiple context-specific sub-modules are merged with a single shared backbone model. This combining approach allows the system to achieve the precision benefits of context-specific modeling while reducing overall complexity by sharing common processing resources across different acoustic contexts through the unified shared layers.

Inventive Principle:
Principle #5Merging (Combining)

3Device complexity

If a single acoustic model is trained on diverse acoustic contexts, then the device complexity is reduced, but the measurement precision deteriorates due to inability to accurately process signals outside the specific training context

Engineering Contradiction:
Improvemodel structureVSAvoidrobustness across acoustic scenarios
Core Design Contradiction:
Device complexityVSMeasurement precision

Solution Approach 1:

The model is segmented into shared layers trained on diverse acoustic contexts and context-specific sub-modules trained on particular scenarios. This segmentation enables the system to maintain simplicity through the shared backbone while achieving robustness across different acoustic scenarios through specialized sub-modules that can be selected or activated based on the input context.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

Different parts of the model have different quality specializations: the shared layers provide general-purpose feature extraction with broad applicability, while context-specific sub-modules provide localized optimization for particular acoustic scenarios. This local quality approach ensures high precision for specific contexts while maintaining overall system simplicity through the shared foundation.

Inventive Principle:
Principle #3Local quality

Data Source

PatentEP3424044B1Modular deep learning model
Publication Date: 2024.08.07 MICROSOFT TECHNOLOGY LICENSING LLC
  • EP3424044B1 patent drawingFigure 1
  • EP3424044B1 patent drawingFigure 2
  • EP3424044B1 patent drawingFigure 3

AI summary

The technology described herein uses a modular model to process speech. A deep learning based acoustic model comprises a stack of different types of neural network layers. The sub-modules of a deep learning based acoustic model can be used to represent distinct non-phonetic acoustic factors, such as accent origins (e.g. native, non-native), speech channels (e.g. mobile, bluetooth, desktop etc.), speech application scenario (e.g. voice search, short message dictation etc.), and speaker variation (e.g. individual speakers or clustered speakers), etc. The technology described herein uses certain sub-modules in a first context and a second group of sub-modules in a second context.