Signal Generation Processing with Parallel Noise-Level Submodels
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing neural vocoder models like WaveGrad and DiffWave face challenges in achieving high-quality audio synthesis while maintaining processing speed, particularly when trained on single-speaker data, and there is a need for improved signal generation processing that maintains quality and speed for both audio and image signals.
Innovation Solution
A signal generation processing device utilizing multiple sub-model units, each trained with different noise levels, allowing parallel training and prediction processing, with a control unit managing the order and selection of sub-models based on noise schedules to optimize processing speed and accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If WaveGrad or DiffWave models are used for audio synthesis, then the number of model parameters is reduced, but the quality of synthesized audio waveform signals deteriorates
Solution Approach 1:
The patent divides a single vocoder model into multiple sub-models (first sub-model to N-th sub-model), where each sub-model processes audio synthesis at different noise levels. This segmentation allows the system to achieve high-quality audio synthesis comparable to WaveGlow while maintaining the parameter efficiency of diffusion stochastic models, as each sub-model can be trained independently on specific noise level ranges.
2Manufacturing precision
If WaveGlow is used for audio synthesis, then the quality of synthesized audio waveform signals is improved, but the time required for training processing increases
Solution Approach 1:
The patent segments the training process into multiple independent sub-models, each handling specific noise level ranges. This enables parallel training of multiple sub-models simultaneously, significantly reducing total training time compared to training a single large WaveGlow model sequentially, while achieving comparable audio quality through the collective output of all sub-models.
Solution Approach 2:
The patent performs preliminary training of multiple sub-models in parallel on different noise level ranges before actual audio synthesis. This preliminary distributed training allows the system to achieve high-quality audio synthesis without requiring the extensive training time needed by WaveGlow, as the computational burden is distributed across multiple independently trained sub-models.
3Productivity
If multiple sub-model units are used for signal generation, then the processing speed is improved through parallel training, but the device complexity increases
Solution Approach 1:
The patent designs multiple sub-models with identical or similar architectures that can process different noise level ranges. This universality means each sub-model unit serves the same function (audio synthesis) but operates on different input conditions, allowing for parallel processing and improved productivity while keeping individual unit complexity manageable and reuse possible.
Data Source
AI summary
Provided is a signal generation processing device that achieves audio synthesis processing or image signal generation processing capable of obtaining high-quality audio signals or image signals while maintaining the speed of audio synthesis processing or image signal generation processing. In the signal generation processing device, the first sub-model unit to the N-th sub-model unit each performs training processing for training models included in the first sub-model unit to the Nth sub-model unit using noise levels included in different noise level ranges to obtain trained models. In other words, the signal generation processing device performs processing for each sub-model unit in parallel, thus allowing for performing the training processing at high speed. Further, during prediction processing, the signal generation processing device appropriately selects the sub-model units to be used and performs processing with the selected sub-models, thus allowing for performing audio synthesis processing and image generation processing with high accuracy.