Speech Enhancement via Deep Dictionary Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional speech noise cancellation algorithms are ineffective in complex environments due to their assumption that noise signals change slowly, leading to suboptimal performance when noise conditions differ from the training set, and require extensive clean and noise signal collection for neural network training.
Innovation Solution
A deep dictionary is created using clean speech signals through deep decomposition and neural network training, allowing for better representation and enhancement of noisy speech, improving generalization ability and efficiency by using only clean speech for training.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If conventional speech noise cancellation algorithms are used based on statistical analysis, then the algorithm is simple to implement, but it cannot provide expected effect when noise conditions differ from training assumptions
Solution Approach 1:
The patent transforms the speech enhancement problem from direct noisy speech processing to a two-stage process: first decomposing clean speech into sparse coefficients and basis matrices, then reconstructing enhanced speech by combining these with noisy speech features. This parameter transformation enables the system to adapt to varying noise conditions while maintaining implementation feasibility through established signal processing operations.
Solution Approach 2:
The patent introduces sparse coefficients and basis matrices as intermediary representations between the input noisy speech and the output enhanced speech. These intermediaries capture the essential speech structure independent of noise, allowing the system to maintain reliability across different noise environments while keeping the overall algorithm structure manageable.
2Reliability
If neural network training with extensive clean and noise signal collection is performed, then noise cancellation performance can be improved, but the data collection and training complexity increases
Solution Approach 1:
The patent extracts the essential speech characteristics by decomposing clean speech into sparse coefficients and basis matrices, separating the speech content from noise. This extraction process eliminates the need for extensive noisy speech data collection, as the model learns speech representations from clean speech alone, thereby reducing training complexity while maintaining performance.
Solution Approach 2:
The patent segments the speech enhancement task into distinct components: clean speech decomposition into sparse representations, noisy speech feature extraction, and final reconstruction. This segmentation allows each component to be optimized independently, reducing overall system complexity while improving noise cancellation performance through specialized processing at each stage.
3Loss of information
If deep decomposition of clean speech is performed to obtain sparse matrix and base matrices, then deeper features of clean speech are learned, but the processing time and computational complexity increase
Solution Approach 1:
The patent performs deep decomposition of clean speech into sparse coefficients and basis matrices in advance during the training phase. These pre-computed representations are then reused during the enhancement phase, allowing rapid processing of noisy speech without repeating the computationally intensive decomposition. This preliminary action preserves speech feature quality while reducing real-time processing time.
Solution Approach 2:
The patent applies deep decomposition selectively to clean speech data during training, rather than processing every input signal through the full decomposition pipeline. During actual enhancement, the system uses the pre-learned basis matrices to efficiently process noisy speech, performing only the necessary reconstruction operations. This partial application of decomposition maintains feature representation quality while minimizing computational overhead.
Data Source
AI summary
Embodiments of this disclosure provide a speech enhancement method and apparatus, an electronic device, and a computer-readable storage medium. The method includes: obtaining a clean speech sample; decomposing the clean speech sample to obtain a first sparse matrix and m base matrices, values in the first sparse matrix being all positive numbers, and m being a positive integer greater than 1; obtaining, according to the first sparse matrix and a weight matrix of a target neural network, state vectors of neurons in a visible layer of the target neural network; and updating the weight matrix according to the state vectors of the neurons in the visible layer and the clean speech sample, to obtain a deep dictionary used for speech enhancement.