Multi-Encoder Information Processing for Selective Music Generation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing information processing systems struggle to generate songs that match user-designated features, such as changing a melody while maintaining a certain rhythm, due to difficulties in selectively learning specific features of content.
Innovation Solution
An information processing apparatus and method that utilizes multiple encoders to separately extract feature quantities from content, allowing for the generation of a learned model that can selectively learn and generate new content based on user-specified features by using a variational autoencoder (VAE) and generative adversarial networks (GANs).
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a single encoder is used to learn features from content, then the learning process is simple, but it is difficult to selectively learn specific features (e.g., rhythm, melody) designated by a user
Solution Approach 1:
The patent divides the feature extraction process into multiple independent encoders, where each encoder is responsible for extracting specific feature quantities (e.g., rhythm, melody, harmony) from the input content. This segmentation allows the system to selectively learn and process individual features independently, enabling users to designate which features to preserve or modify while generating new content.
2Adaptability or versatility
If multiple encoders are introduced to extract specific feature quantities, then selective feature learning becomes possible, but the device complexity increases
Solution Approach 1:
The patent extracts specific feature quantities (such as rhythm, melody, and harmony) from the overall content using dedicated encoders. Each encoder is designed to extract a particular type of feature, allowing the system to isolate and process individual features independently. This extraction approach enables flexible feature selection without requiring the entire content to be processed uniformly.
3Adaptability or versatility
If all features of content are learned together, then the learning process is straightforward, but it is impossible to generate variations that preserve only specific features while changing others
Solution Approach 1:
The patent segments the feature learning process into multiple independent encoders, each responsible for a specific feature type. This segmentation enables the system to selectively preserve or modify individual features when generating content variations, allowing users to control which features remain constant and which can be changed or transformed.
Solution Approach 2:
The patent applies different learning characteristics to different feature types through specialized encoders. Each encoder is optimized to capture the unique properties of its designated feature (e.g., rhythm encoders focus on temporal patterns, melody encoders focus on pitch sequences), enabling precise control over which features are preserved and which are transformed in the generated content.
Data Source
Figure 1
Figure 2~3
Figure 4
AI summary
An information processing apparatus 100 according to the present disclosure includes an extraction unit 131 that extracts first data from an element constituting first content, and a model generation unit 132 that generates a learned model that has a first encoder 50 that calculates a first feature quantity which is a feature quantity of first content, and a second encoder 55 that calculates a second feature quantity which is a feature quantity of the extracted first data.