Two-Stage Neural Audio Processing for Bit-Rate and Quality Tradeoffs

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing neural network-based audio signal processing systems face impractical complexity due to the need for large amounts of training data and computational resources, and vector quantization methods introduce a trade-off between complexity reduction and quality, limiting performance.

Innovation Solution

A two-stage neural network architecture is employed, where the first network generates a bit-rate reduced representation associated with a pre-defined audio encoding process, and the second network enhances this representation to produce the final output, allowing separate training and optimizing the trade-off between bit-rate reduction and distortion.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If a single neural network system is used for audio signal processing, then the system can be trained end-to-end, but the computational complexity and training data requirements become impractically large

Engineering Contradiction:
Improvesignal processing qualityVSAvoidneural network complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent divides the single complex neural network into multiple specialized sub-networks (encoder, quantizer, decoder, noise predictor, etc.), each performing a specific function in the audio signal processing pipeline. This segmentation reduces the complexity of individual networks while maintaining overall processing quality through coordinated operation of all components.

Inventive Principle:
Principle #1Segmentation

2Device complexity

If vector quantization is applied to reduce complexity, then computational resources are reduced, but information relevant to solving the processing problem is removed, limiting achievable performance

Engineering Contradiction:
Improvecomputational complexityVSAvoidsignal information loss
Core Design Contradiction:
Device complexityVSLoss of information

Solution Approach 1:

The patent changes the parameter representation of audio signals by transforming them into a latent space and applying quantization in this transformed domain rather than directly to the original signal. This allows for more efficient representation with reduced information loss, as the latent space captures essential signal characteristics in a compressed form suitable for quantization.

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent introduces an intermediate latent representation layer between the encoder and decoder, where quantization is applied. This intermediary latent space serves as a bridge that allows compression while preserving essential information through learned transformations, rather than directly quantizing the original audio signal.

Inventive Principle:
Principle #24Intermediary (Mediator)

3Device complexity

If vector quantization is applied directly to the signal, then complexity is reduced, but the trade-off between complexity reduction and attainable quality is difficult to optimize and limits performance

Engineering Contradiction:
Improvemodel complexityVSAvoidsignal processing quality
Core Design Contradiction:
Device complexityVSManufacturing precision

Solution Approach 1:

The patent implements dynamic quantization strategies where the quantization process adapts to the characteristics of the input signal. The noise predictor and variable rate quantizer adjust quantization parameters based on signal properties, allowing the system to maintain high quality for complex signals while achieving compression for simpler signals, thus optimizing the complexity-quality trade-off dynamically.

Inventive Principle:
Principle #15Dynamics

Data Source

PatentEP4558987B1Neural network based signal processing
Publication Date: 2026.03.25 DOLBY INTERNATIONAL AB
  • EP4558987B1 patent drawingFigure 1~2
  • EP4558987B1 patent drawingFigure 3
  • EP4558987B1 patent drawingFigure 4~5

AI summary

A method for processing an input audio signal, comprising conditioning a first neural network system with a representation of the input audio signal to predict a bit-rate reduced representation of a processed input audio signal, the first neural network system being trained to generate a bit-rate reduced representation of a processed version of a given audio signal, wherein the bit-rate reduced representation has a format associated with a pre-defined audio encoding process, conditioning a second neural network system with the bit-rate reduced representation to predict an enhanced representation of the processed audio signal, the second neural network system being trained to generate an enhanced representation of a given a bit-rate reduced audio representation, wherein the bit-rate reduced representation has a format associated with the pre-defined audio encoding process, and transforming the enhanced representation of the processed audio signal into an output audio signal.