Deep Generative Speech Enhancement for Coded Audio Distortion

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing speech coding technologies suffer from quality issues due to decreasing bitrates, leading to distorted and contaminated speech signals that are difficult to restore effectively.

Innovation Solution

A system utilizing self-supervised deep learning models to generate feature vectors from contaminated, coded speech, followed by a generative deep learning model to produce high-quality clean speech signals, leveraging robust features from low-level audio properties to high-level speaker characteristics.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Loss of energy

If speech coding bitrate is decreased to enable low-cost mobile and internet communication, then communication cost is reduced, but speech quality deteriorates with various quality issues

Engineering Contradiction:
Improvecommunication costVSAvoidspeech quality
Core Design Contradiction:
Loss of energyVSManufacturing precision

Solution Approach 1:

The patent introduces a denoising autoencoder with linear regression decoder as an intermediary component between the coded speech and the final output. This intermediary model processes the noisy coded speech through learned representations to reconstruct clean speech, effectively mediating the quality degradation caused by low bitrate coding while maintaining communication cost efficiency.

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

The patent changes the parameters of the speech signal by transforming it through a deep learning model that learns optimal representations. The model adjusts internal parameters (weights and biases) during training to map coded speech to clean speech, effectively changing the signal characteristics to improve quality without increasing bitrate.

Inventive Principle:
Principle #35Parameter changes

2Device complexity

If conventional audio processing techniques are used to restore clean speech from coded audio data, then processing simplicity is maintained, but speech restoration quality is insufficient

Engineering Contradiction:
Improveprocessing simplicityVSAvoidspeech restoration quality
Core Design Contradiction:
Device complexityVSManufacturing precision

Solution Approach 1:

The patent replaces conventional mechanical audio processing techniques with a data-driven deep learning approach. Instead of using traditional signal processing algorithms, the system employs a trained denoising autoencoder model that automatically learns the transformation from coded to clean speech, achieving superior restoration quality.

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

Solution Approach 2:

The patent performs preliminary training of the denoising autoencoder model using training data before actual speech restoration. This preliminary action prepares the model with learned representations and parameters, enabling it to effectively restore speech quality when processing actual coded audio data in application scenarios.

Inventive Principle:
Principle #10Preliminary action

3Manufacturing precision

If deep learning models are used to generate improved speech data, then speech quality is enhanced, but computational complexity increases

Engineering Contradiction:
Improvespeech qualityVSAvoidcomputational complexity
Core Design Contradiction:
Manufacturing precisionVSDevice complexity

Solution Approach 1:

The patent segments the deep learning model into distinct functional components: an encoder that extracts features from coded speech, a denoising autoencoder that processes hidden representations, and a linear regression decoder that reconstructs clean speech. This segmentation allows for optimized computation at each stage and enables real-time processing through efficient inference.

Inventive Principle:
Principle #1Segmentation

Data Source

PatentEP4483360B1Coded speech enhancement based on deep generative model
Publication Date: 2025.12.03 DOLBY LABORATORIES LICENSING CORP
  • EP4483360B1 patent drawingFigure 1~2
  • EP4483360B1 patent drawingFigure 3~4
  • EP4483360B1 patent drawingFigure 5

AI summary

A system for generating enhanced speech data using robust audio features is disclosed. In some embodiments, a system is programmed to use a self-supervised deep learning model to generate a set of feature vectors from given audio data that contains contaminated speech and is coded. The system is further programmed to use a generative deep learning model to create improved audio data corresponding to clean speech from the set of feature vectors.