Dual-Path Neural Network for Speech Signal Encoding
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing encoding methods struggle to efficiently encode both periodic and non-periodic components of sequential signals like speech and audio, leading to suboptimal compression efficiency and reconstruction quality.
Innovation Solution
A dual-path neural network combined with a gating neural network is used to separately model periodic and non-periodic components of sequential signals, enhancing compression efficiency and reconstruction quality.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If a feedforward-type autoencoder scheme is used to encode sequential signals, then the encoding structure is simple, but the temporal structure and periodicity of speech or audio signals cannot be effectively modeled
Solution Approach 1:
The patent segments the encoding process into two distinct paths: a feedforward neural network path for capturing non-periodic components and a recurrent neural network path for capturing temporal structure and periodicity. This segmentation allows each path to be optimized for its specific function while working together to solve the overall encoding problem.
Solution Approach 2:
The patent merges the feedforward neural network and recurrent neural network into a unified dual-path architecture. The outputs from both paths are combined through a gating mechanism to produce the final encoded representation, allowing the system to leverage both simple feedforward processing and complex temporal modeling capabilities.
2Reliability
If a recurrent-type autoencoder scheme is used to encode sequential signals, then the temporal structure can be effectively modeled, but the encoding efficiency for non-periodic components deteriorates
Solution Approach 1:
The patent segments the signal processing into two parallel paths, allowing non-periodic components to be processed efficiently by the feedforward network while periodic components are handled by the recurrent network. This segmentation prevents the recurrent network from processing all data, thereby improving overall encoding efficiency.
Solution Approach 2:
The patent applies partial action by having the recurrent neural network focus only on processing the periodic components of the signal, rather than processing the entire signal. The gating mechanism selectively routes appropriate portions of the data to each network path based on the characteristics of the input signal.
3Device complexity
If a single neural network is used to encode both periodic and non-periodic components, then the device complexity is low, but the reconstruction quality deteriorates
Solution Approach 1:
The patent segments the encoding task into two specialized sub-tasks handled by separate neural network paths: one for periodic components and one for non-periodic components. This segmentation allows each network to be optimized for its specific task, improving overall reconstruction quality compared to a single general-purpose network.
Solution Approach 2:
The patent applies local quality by giving different structural characteristics to different parts of the encoding system. The feedforward path uses simple layered structures suitable for non-periodic patterns, while the recurrent path uses looped structures suitable for temporal dependencies, allowing each part to have the quality appropriate for its specific function.
4Manufacturing precision
If separate encoding paths are used for periodic and non-periodic components, then the reconstruction quality is improved, but the device complexity increases
Solution Approach 1:
The patent merges the two separate encoding paths through a gating mechanism that combines their outputs into a unified representation. This merging approach maintains the benefits of specialized processing while presenting a cohesive interface to the rest of the system, thereby managing complexity more effectively.
Solution Approach 2:
The gating mechanism serves multiple functions: it combines outputs from both paths, selectively weights their contributions based on signal characteristics, and produces a unified encoded representation. This multi-functionality reduces the need for additional separate components, thereby managing overall system complexity.
Data Source
AI summary
An encoding method, a decoding method, an encoder for performing the encoding method, and a decoder for performing the decoding method are provided. The encoding method includes outputting LP coefficients bitstream and a residual signal by performing an LP analysis on an input signal, outputting a first latent signal obtained by encoding a periodic component of the residual signal, a second latent signal obtained by encoding a non-periodic component of the residual signal, and a weight vector for each of the first latent signal and the second latent signal, using a first neural network module, and outputting a first bitstream obtained by quantizing the first latent signal, a second bitstream obtained by quantizing the second latent signal, and a weight bitstream obtained by quantizing the weight vector, using a quantization module.