Frequency Component Vector Generation via Neural Network Segmentation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

The existing WaveNet and NPSS methods for generating sound sample sequences face limitations due to the need for simultaneous probability distributions of all dimensions, which results in a large number of probability distributions and parameters, leading to high calculation resources and potential training issues, especially with insufficient training data for the number of dimensions.

Innovation Solution

A method involving a main neural network and sub-neural networks that generate frequency component vectors by processing data in unit steps, where each sub-neural network receives intermediate data and either a component value or intermediate value from another frequency band to produce component values, reducing the overall calculation load and allowing for conditional probability distributions based on dependencies among dimensions.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Speed

If conventional Fourier transform methods are used to extract frequency components, then frequency analysis can be performed, but the computational complexity increases significantly and processing time is excessive for real-time applications

Engineering Contradiction:
Improveprocessing speedVSAvoidfrequency component extraction accuracy
Core Design Contradiction:
SpeedVSMeasurement precision

Solution Approach 1:

The patent segments the time-series data into multiple blocks and processes each block separately using FFT. This divides the large-scale computationally intensive Fourier transform into smaller, manageable segments that can be processed quickly and in parallel, significantly reducing overall processing time while maintaining frequency analysis capability

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent performs preliminary processing by calculating the mean and standard deviation of each segmented block before combining them. This preliminary action prepares the data in a form that enables efficient combination later, reducing the computational burden during the final frequency component extraction stage

Inventive Principle:
Principle #10Preliminary action

2Productivity

If simple averaging of segmented frequency components is performed, then processing time is reduced, but the accuracy of frequency component extraction deteriorates

Engineering Contradiction:
Improveprocessing efficiencyVSAvoidfrequency component accuracy
Core Design Contradiction:
ProductivityVSMeasurement precision

Solution Approach 1:

The patent changes the combination parameter from simple arithmetic mean to a weighted combination method that considers the standard deviation of each block. This parameter change allows the system to give more weight to blocks with lower variability, improving the accuracy of the combined frequency components while maintaining processing efficiency

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent incorporates feedback by calculating the standard deviation of each block and using this information to adjust the weighting in the combination process. This feedback mechanism ensures that blocks with more reliable frequency components (lower standard deviation) contribute more to the final result, improving overall accuracy

Inventive Principle:
Principle #23Feedback

3Reliability

If all frequency components are processed to ensure comprehensive analysis, then analysis completeness is improved, but the computational load and processing time increase excessively

Engineering Contradiction:
Improveanalysis completenessVSAvoidcomputational complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent extracts only the essential frequency components from each segmented block rather than processing all possible frequency information. By taking out and processing only the necessary frequency components, the system maintains comprehensive analysis capability while significantly reducing computational complexity and processing requirements

Inventive Principle:
Principle #2Taking out (Extraction)

Data Source

PatentEP3836027A1Method and device for generating frequency component vector of time-series data
Publication Date: 2021.06.16 YAMAHA CORP
  • EP3836027A1 patent drawingFigure 1~2
  • EP3836027A1 patent drawingFigure 3
  • EP3836027A1 patent drawingFigure 4

AI summary

This method generates a frequency component vector of time-series data in each unit step of a plurality of unit steps, by executing a first process and a second process in each unit step. In the first process, the first data is received and processed using a first neural network to generate intermediate data as an alternative representation of the first data. In the second process, a plurality of component values corresponding to a plurality of frequency bands are generated based on the generated intermediate data. The second process includes a plurality of unit processes corresponding to the plurality of frequency bands. Each of the plurality of unit processes includes an inference by a second neural network supplied with the intermediate data to generate a corresponding one of the plurality of component values. A first unit process corresponding to a first frequency band among the plurality of unit processes supplies, the intermediate data and either a second component value generated by a second unit process corresponding to a second frequency band different from the first frequency band or an intermediate value in the second unit process to the second neural network to generate a first component value.