Speech Recognition Neural Network Attention Mechanism

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing speech recognition technologies face challenges in accurately recognizing speech in noisy environments and distinguishing between multiple speakers, as they lack the ability to selectively focus on specific spectral areas and adaptively remove noise.

Innovation Solution

A speech recognizing model implemented through a neural network that dynamically applies attention weights to the input speech signal, based on previous layer outputs and context values, to emphasize important frequency areas and reduce noise interference, thereby enhancing recognition performance.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If traditional speech recognition models process the entire speech signal uniformly, then the model structure is simple, but the recognition accuracy in noisy environments deteriorates

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

Solution Approach 1:

The speech signal is segmented into multiple frequency bands through spectral decomposition. The attention mechanism divides the spectral representation into distinct regions, allowing the model to process different frequency components separately rather than uniformly, thereby improving noise robustness while maintaining manageable complexity through localized processing.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

Different attention weights are applied to different spectral regions based on their local characteristics. The model learns to apply varying degrees of attention to different frequency bands, emphasizing speech-relevant regions while suppressing noise-dominated regions, thus achieving high recognition accuracy through localized adaptive processing.

Inventive Principle:
Principle #3Local quality

2Adaptability or versatility

If the model processes all frequency components with equal weight, then the processing is straightforward, but the ability to distinguish speakers and remove noise deteriorates

Engineering Contradiction:
Improvenoise adaptation capabilityVSAvoidattention mechanism complexity
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The attention weights are dynamically adjusted based on the input speech signal characteristics and noise conditions. The model learns to adaptively reweight spectral components in real-time, allowing it to respond to varying noise environments and speaker characteristics, thereby achieving high adaptability through dynamic rather than static processing.

Inventive Principle:
Principle #15Dynamics

Solution Approach 2:

The attention mechanism incorporates feedback loops where the model's predictions and intermediate representations inform the attention weight calculation. This feedback enables the model to iteratively refine its focus on relevant spectral regions, improving noise adaptation and speaker differentiation through continuous adjustment based on performance feedback.

Inventive Principle:
Principle #23Feedback

3Reliability

If the model focuses on specific spectral areas, then the noise reduction performance improves, but the computational complexity increases

Engineering Contradiction:
Improvespeaker identification reliabilityVSAvoidcomputational energy consumption
Core Design Contradiction:
ReliabilityVSUse of energy by moving object

Solution Approach 1:

The attention mechanism extracts and isolates speaker-specific spectral features from the mixed speech signal by selectively focusing on relevant frequency regions. By taking out and processing only the most informative spectral components rather than the entire spectrum, the model achieves reliable speaker identification with reduced computational energy consumption compared to uniform full-spectrum processing.

Inventive Principle:
Principle #2Taking out (Extraction)

Data Source

PatentEP3346462B1Speech recognizing method and apparatus
Publication Date: 2022.06.08 SAMSUNG ELECTRONICS CO LTD
  • EP3346462B1 patent drawingFigure 1~2
  • EP3346462B1 patent drawingFigure 3
  • EP3346462B1 patent drawingFigure 4

AI summary

A speech recognizing method and apparatus is provided. A speech recognizing method, implementing a speech recognizing model neural network for recognition of a speech, includes determining an attention weight based on an output value output by at least one layer of the speech recognizing model neural network at a previous time of the recognition of the speech, applying the determined attention weight to a speech signal corresponding to a current time of the recognition of the speech, and recognizing the speech signal to which the attention weight is applied, using the speech recognizing model neural network.