Dual-Path Neural Network for Speech Signal Encoding
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional encoding and decoding methods are inefficient for encoding periodic and aperiodic components of speech and audio signals, leading to suboptimal compression and quality in sequential signal processing.
Innovation Solution
A dual-path neural network structure is employed, utilizing recurrent neural networks for periodic component modeling and feed-forward neural networks for aperiodic component modeling, with spectral whitening through linear prediction analysis to enhance encoding and decoding efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If a feed-forward autoencoder method is used for encoding, then encoding efficiency for non-sequential signals is improved, but encoding efficiency for sequential signals containing periodicity deteriorates
Solution Approach 1:
The patent segments the sequential signal into periodic and aperiodic components, applying different encoding strategies to each. The periodic component is modeled using one neural network path while the aperiodic component uses another path, allowing each component to be optimized independently for its specific characteristics.
Solution Approach 2:
The patent dynamically adapts the encoding approach by selectively applying different neural network architectures based on the signal characteristics. For periodic components, a recurrent neural network structure is used, while for aperiodic components, a feed-forward structure is applied, making the system flexible and adaptive to different signal types.
2Reliability
If a recurrent autoencoder method is used for encoding sequential signals, then temporal structure modeling is improved, but encoding efficiency for aperiodic components deteriorates
Solution Approach 1:
The patent divides the signal processing into separate pathways: one dedicated to periodic components using recurrent neural networks for temporal modeling, and another for aperiodic components using feed-forward networks for efficient encoding, allowing each path to be optimized for its specific function.
Solution Approach 2:
Different neural network structures are applied to different parts of the signal based on their local characteristics. Periodic portions receive recurrent processing suitable for temporal patterns, while aperiodic portions receive feed-forward processing optimized for non-repetitive structures, achieving local optimization throughout the signal.
3Device complexity
If traditional encoding methods are used, then device complexity is reduced, but compression efficiency and quality deteriorate
Solution Approach 1:
The encoding structure is segmented into multiple specialized neural network paths, each handling specific signal components. This segmentation allows the system to achieve high compression efficiency by targeting different signal characteristics with appropriate processing, while maintaining manageable complexity through modular architecture.
Solution Approach 2:
The patent creates a universal encoding framework that can handle both periodic and aperiodic components within a single system. The multi-path neural network structure provides multi-functionality, accommodating different signal types and characteristics without requiring separate dedicated systems for each case.
4Device complexity
If traditional encoding methods are used, then device complexity is reduced, but encoding quality deteriorates
Solution Approach 1:
The encoding system is segmented into specialized pathways that target specific signal components, allowing each path to achieve high encoding quality for its designated function. This segmentation enables precise modeling of periodic and aperiodic components separately, improving overall encoding quality while maintaining modular complexity.
Solution Approach 2:
The patent changes the parameters and architecture of the neural networks to optimize encoding quality. By adjusting network structures, activation functions, and processing parameters within each path, the system achieves high-fidelity encoding while managing complexity through parameter optimization rather than structural bloat.
Data Source
AI summary
Provided is an encoding method according to various example embodiments and an encoder performing the method. The encoding method includes outputting a linear prediction (LP) coefficients bitstream and a residual signal by performing a linear prediction analysis on an input signal, outputting a first latent signal obtained by encoding a periodic component of the residual signal, using a first neural network module, outputting a first bitstream obtained by quantizing the first latent signal, using a quantization module, outputting a second latent signal obtained by encoding an aperiodic component of the residual signal, using the first neural network module, and outputting a second bitstream obtained by quantizing the second latent signal, using the quantization module, wherein the aperiodic component of the residual signal is calculated based on a periodic component of the residual signal decoded from the quantized first latent signal output by de-quantizing the first bitstream.