LSTM Speech Enhancement Parameter Extraction

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing speech enhancement technologies require high storage capacity to store noise reduction models for different speakers, which is inefficient and costly.

Innovation Solution

A method using a long short-term memory (LSTM) model to perform speech enhancement by obtaining and updating speech enhancement parameters for individual speakers, reducing the need for multiple models by storing only parameters and continuously training on new data.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Manufacturing precision

If noise reduction models are stored for different speakers, then speech enhancement quality is improved, but storage capacity requirements increase

Engineering Contradiction:
Improvespeech enhancement qualityVSAvoidstorage capacity
Core Design Contradiction:
Manufacturing precisionVSQuantity of substance

Solution Approach 1:

The patent extracts only the essential speech enhancement parameters from complete noise reduction models. Instead of storing entire models for each speaker, it identifies and stores only the critical parameters that capture individual speech characteristics, significantly reducing storage requirements while maintaining enhancement quality.

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The patent transforms the representation of speaker-specific noise reduction capabilities from complete models to condensed parameter sets. By changing the data structure from full models to essential parameters, it achieves both quality preservation and storage reduction.

Inventive Principle:
Principle #35Parameter changes

2Manufacturing precision

If adaptive training is performed on entire noise reduction models, then speaker-specific enhancement is improved, but training time and computational resources increase

Engineering Contradiction:
Improvespeaker-specific enhancementVSAvoidtraining time
Core Design Contradiction:
Manufacturing precisionVSLoss of time

Solution Approach 1:

The patent extracts only the essential parameters that need to be trained for each speaker, rather than performing adaptive training on entire noise reduction models. This selective parameter training dramatically reduces computational resources and training time while maintaining speaker-specific enhancement quality.

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The patent applies partial training by focusing only on the essential parameters that capture speaker characteristics, rather than training the complete model. This partial action approach achieves sufficient speaker adaptation with significantly reduced computational overhead.

Inventive Principle:
Principle #16Partial or excessive action

Data Source

PatentEP3920183B1Speech data processing method and apparatus, electronic device and readable storage medium
Publication Date: 2025.06.25 TENCENT TECHNOLOGY (SHENZHEN) CO LTD
  • EP3920183B1 patent drawingFigure 1A~1B
  • EP3920183B1 patent drawingFigure 2~3
  • EP3920183B1 patent drawingFigure 4~5

AI summary

A speech data processing method and apparatus, an electronic device and a computer-readable storage medium. The method comprises: receiving first speech data sent by a sending party, and acquiring corresponding speech enhancement parameters (S101); performing, on the basis of the acquired speech enhancement parameters, speech enhancement processing on the first speech data to obtain first speech enhancement data, and determining first speech enhancement parameters on the basis of the first speech enhancement data (S102); and sending the first speech enhancement data to a receiving party, and updating the acquired speech enhancement parameters with the first speech enhancement parameters so as to obtain updated speech enhancement parameters for use in performing, when second speech data sent by the sending party is received, speech enhancement processing on the second speech data on the basis of the updated speech enhancement parameters (S103).