Compound Speech Models for Speaker Adaptation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current speech recognition systems face challenges in accommodating multiple speakers with different characteristics, such as gender, age, and accent, as purely speaker-dependent models require extensive training and purely speaker-independent models are less accurate, leading to increased memory usage and complexity.

Innovation Solution

The development of compound models that incorporate both speaker-independent and speaker-dependent portions, allowing for efficient clustering and adaptation, reducing memory consumption and enabling robust speech processing across various speakers without the need for separate modules.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If speaker-dependent models are used to improve speech recognition accuracy for specific speakers, then recognition accuracy is improved, but memory usage and model complexity increase significantly

Engineering Contradiction:
Improvespeech recognition accuracyVSAvoidmemory usage
Core Design Contradiction:
Measurement precisionVSQuantity of substance

Solution Approach 1:

The speech recognition system is segmented into two distinct components: speaker-independent models that handle general speech patterns and speaker-dependent models that handle individual speaker characteristics. This segmentation allows the system to load only necessary speaker-specific models into memory rather than storing complete speaker-dependent models for all possible speakers, thereby reducing memory usage while maintaining accuracy for recognized speakers.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The speaker-independent model serves as a universal base that can be applied to any speaker, handling general speech recognition tasks. This universal component reduces the need for separate complete models for each speaker, as the speaker-dependent adaptations can be applied on top of the shared speaker-independent framework, reducing overall memory requirements.

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

2Quantity of substance

If speaker-independent models are used to reduce memory usage and simplify the system, then memory consumption is reduced, but speech recognition accuracy deteriorates

Engineering Contradiction:
Improvememory usageVSAvoidspeech recognition accuracy
Core Design Contradiction:
Quantity of substanceVSMeasurement precision

Solution Approach 1:

The system dynamically adapts the speaker-independent model by applying speaker-dependent transformations when speaker characteristics are identified. This dynamic adaptation allows the system to maintain high accuracy for known speakers while keeping the base speaker-independent model lightweight for memory efficiency. The speaker-dependent parameters are applied on-demand rather than being permanently stored.

Inventive Principle:
Principle #15Dynamics

Solution Approach 2:

The system changes parameters of the speaker-independent model based on detected speaker characteristics. By modifying model parameters dynamically according to speaker identity rather than maintaining separate complete models, the system achieves speaker-specific accuracy while keeping memory usage low through parameter adaptation rather than structural duplication.

Inventive Principle:
Principle #35Parameter changes

3Measurement precision

If separate speaker-dependent models are trained for each speaker to achieve high accuracy, then speech recognition accuracy is improved, but device complexity and training requirements increase

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

Solution Approach 1:

The model architecture is segmented into a shared speaker-independent component and speaker-dependent adaptation layers. This segmentation simplifies the overall architecture by avoiding the need for completely separate models for each speaker, reducing device complexity while maintaining the ability to achieve high accuracy through targeted speaker-specific adaptations.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The system merges the speaker-independent model with speaker-dependent adaptations into a unified framework. This combination reduces device complexity by consolidating multiple speaker-specific models into a single adaptable system, while still achieving high accuracy through the integrated speaker-dependent components that adjust the shared base model for individual speakers.

Inventive Principle:
Principle #5Merging (Combining)

Data Source

PatentEP2389672B1Method, apparatus and computer program product for providing compound models for speech recognition adaptation
Publication Date: 2018.08.08 NOKIA TECHNOLOGIES OY
  • EP2389672B1 patent drawingFigure 1
  • EP2389672B1 patent drawingFigure 2
  • EP2389672B1 patent drawingFigure 3

AI summary

An apparatus for providing compound models for speech recognition adaptation includes a processor. The processor may be configured to receive a speech signal corresponding to a particular speaker, select a cluster model including both a speaker independent portion and a speaker dependent portion based at least in part on a characteristic of speech of the particular speaker, and process the speech signal using the selected cluster model. A corresponding method and computer program product are also provided.