Virtual Noise Signal Generation for Deep Learning Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing noise elimination models trained with artificially generated noisy signals perform poorly when applied to real-world voice signals, and acquiring diverse real-world noise data for training is time-consuming and costly.
Innovation Solution
A data generating apparatus and method that convert real-world noisy and original sound signals into short-time frequency domain spectra, synchronizing them, and using deep neural networks to generate virtual noisy signals similar to real environments for training noise elimination models, and a noise eliminating apparatus that uses these models to remove noise from signals.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If noise elimination models are trained with artificially generated noisy signals, then training data can be obtained easily, but the recognition performance on real-world signals is poor
Solution Approach 1:
The patent creates virtual noisy signals by copying the statistical characteristics and temporal structures of real noise from recorded environmental sounds. Instead of using simple artificial noise addition, the system records actual noise environments (street, office, home, etc.) and uses these recordings to generate virtual noisy training signals that replicate real-world noise patterns, thereby improving model performance while maintaining ease of data generation
Solution Approach 2:
The patent transforms the training data generation process by changing parameters from simple artificial noise addition to complex virtual signal synthesis. The system adjusts parameters such as noise type, signal-to-noise ratio, temporal structure, and spectral characteristics to create diverse virtual noisy signals that match real-world conditions, resolving the contradiction between ease of data generation and recognition accuracy
2Reliability
If large amounts of real environment noisy data are acquired for training, then model performance improves, but the process is time-consuming and costly
Solution Approach 1:
The patent performs preliminary action by recording and storing various types of real environmental noise in advance (street noise, office noise, home noise, etc.). These pre-recorded noise samples are then used to generate virtual noisy signals on-demand during model training, eliminating the need to acquire real noisy data during the training process itself. This preliminary preparation maintains high model performance while significantly reducing training time
Solution Approach 2:
Instead of acquiring large amounts of real noisy data during training, the system copies the essential characteristics of real noise from pre-recorded samples and uses these copies to generate virtual training signals. This copying approach preserves the authenticity of real-world noise patterns while avoiding the time-consuming process of collecting and processing actual noisy recordings for each training iteration
3Adaptability or versatility
If various types of noisy signals are obtained from real environment, then training data diversity improves, but data acquisition difficulty increases
Solution Approach 1:
The patent introduces dynamics by creating a versatile virtual noise generation system that can adaptively produce various types of noisy signals based on predefined noise categories (street, office, home, transportation, etc.). The system dynamically adjusts noise parameters, mixing ratios, and temporal characteristics to generate diverse training signals without requiring physical presence in different environments, thereby maintaining data diversity while simplifying acquisition
Solution Approach 2:
The patent creates a universal noise generation framework that uses a single set of pre-recorded noise samples to generate multiple types of noisy signals for different scenarios. The virtual signal generation system serves multiple functions by adjusting parameters to simulate various noise environments, eliminating the need to separately acquire data from each real-world location while maintaining comprehensive data diversity
Data Source
AI summary
A data generating apparatus for generating noise environment noisy data is disclosed. The data generating apparatus according to the present application comprises a signal conversion unit configured to convert each of a noisy signal obtained in real environment and an original sound signal for the noisy signal into a noisy signal spectrum and an original sound signal spectrum in a short-time frequency domain; and a noisy signal generation training unit configured to train deep neural network to output the noisy signal spectrum corresponding to each short-time using the original sound signal spectrum as an input.