Speech Enhancement Model Training with Frequency Band Compression
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing speech enhancement models, due to their high operation complexity, struggle to meet real-time operation requirements in communication scenarios, leading to communication delays and degraded experiences.
Innovation Solution
A training method that reduces feature dimensionality through frequency band compression and employs cyclic iteration in feature mapping to enhance model training, without adding model structure, resulting in a trained model that reduces operation complexity and meets real-time needs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a speech enhancement model with multiple layers of neural networks (CNN, LSTM, GRU) is used, then speech enhancement capability is improved, but operation complexity increases
Solution Approach 1:
The patent extracts and removes redundant or less important neural network components (such as LSTM or GRU layers) from the speech enhancement model, retaining only the essential CNN layers that provide sufficient speech enhancement capability. This extraction approach reduces operation complexity while preserving the core functionality needed for speech enhancement in real-time communication scenarios.
2Reliability
If a speech enhancement model with high operation complexity is used, then speech enhancement capability is improved, but real-time operation requirement cannot be met
Solution Approach 1:
The patent removes computationally intensive components from the model architecture, keeping only the lightweight CNN layers that can process speech signals quickly. This extraction of essential functionality enables the model to meet real-time operation requirements while maintaining adequate speech enhancement capability.
Solution Approach 2:
The patent adopts a simpler, more lightweight model architecture that prioritizes fast processing over complex feature extraction. This approach uses computationally inexpensive operations that can be executed rapidly, sacrificing some enhancement quality to achieve real-time performance suitable for communication scenarios.
3Reliability
If a speech enhancement model with high operation complexity is used, then speech enhancement capability is improved, but communication delay increases
Solution Approach 1:
The patent extracts and eliminates time-consuming processing stages from the speech enhancement pipeline by removing complex recurrent neural network layers. The simplified CNN-based model processes speech frames more quickly, reducing the overall communication delay while preserving essential noise suppression and speech quality improvement functions.
Data Source
AI summary
This present disclosure relates to a speech enhancement model training method and apparatus, an electronic device, and a storage medium. The method includes: extracting a first audio feature of a to-be-enhanced speech signal through an input layer in each instance of iterative training of an initial speech enhancement model; performing frequency band compression on the first audio feature through a frequency band compression layer, to obtain a dimensionality-reduced second audio feature; performing, through a feature mapping layer, feature mapping on the second audio feature by using a cyclic iteration manner, to obtain a third audio feature, a quantity of output channels of the feature mapping layer increasing progressively in a cyclic iteration process; and inputting the third audio feature to an output layer, to obtain estimated gain information, and performing parameter adjustment on the initial speech enhancement model with reference to true gain information.