GAN Speech Bandwidth Extender for High Frequency Estimation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing deep neural network (DNN)-based speech bandwidth extension methods face limitations in accurately estimating high frequency bands, leading to unpleasant noise due to overestimation and smoothing issues, which affect speech call quality.
Innovation Solution
A generative adversarial network-based method is employed to train a DNN classification model to discriminate between true and fake data, and a DNN generation model is retrained to estimate high frequency bands closely matching the actual wideband signal, maximizing estimation performance without artificial sound interference.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of information
If a DNN generation model is used to estimate high frequency band components, then the speech bandwidth can be extended, but the estimation performance is limited and smoothed values are produced causing unpleasant noise
Solution Approach 1:
The patent introduces a DNN classification model as an intermediary to evaluate and guide the DNN generation model's estimation process. The classification model distinguishes between real and generated high frequency band components, providing feedback to improve the generation model's estimation precision and reduce smoothing artifacts that cause unpleasant noise.
Solution Approach 2:
The patent implements a feedback mechanism where the DNN classification model evaluates the output of the DNN generation model and provides guidance for improvement. This feedback loop enables the generation model to iteratively refine its estimation of high frequency band components, achieving better precision and more natural sound quality.
2Measurement precision
If the DNN classification model is trained to discriminate real from fake data, then the estimation accuracy improves, but the training complexity and computational resources increase
Solution Approach 1:
The patent segments the training process into two distinct phases: first training the DNN classification model to discriminate real from generated data, then using this trained classifier to guide the DNN generation model's training. This segmentation allows each model to be optimized independently, reducing overall training complexity while maintaining high discrimination accuracy.
3Reliability
If the DNN generation model estimates high frequency bands more accurately, then speech call quality improves, but unpleasant noise occurs due to overestimation
Solution Approach 1:
The DNN classification model provides feedback to the DNN generation model by identifying overestimated components. This feedback mechanism enables the generation model to adjust its estimates, improving speech call quality while avoiding the unpleasant noise caused by overestimation of high frequency band components.
Solution Approach 2:
The patent dynamically adjusts the estimation parameters of the DNN generation model based on the classification model's feedback. By changing the estimation parameters adaptively, the system achieves better speech call quality while preventing overestimation that would generate unpleasant noise.
Data Source
AI summary
Proposed are a generative adversarial network-based speech bandwidth extender and extension method. A generative adversarial network-based speech bandwidth extension method, according to an embodiment, comprises the steps of: extracting feature vectors from a narrowband (NB) signal and a wideband (WB) signal of a speech; estimating the feature vector of the wideband signal from the feature vector of the narrowband signal; and learning a deep neural network classification model for discriminating the estimated feature vector of the wideband signal from the actually extracted feature vector of the wideband signal and the actually extracted feature vector of the narrowband signal.