DNN Gain-Mask Training for Audio Noise Suppression Trade-Offs
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing audio enhancement methods fail to efficiently balance noise suppression and signal distortion, requiring expensive and slow expert fine-tuning due to numerous voice and noise characteristics, and are limited by hardware acceleration for real-time performance.
Innovation Solution
Implementing a deep neural network (DNN) model trained with a loss function using clean-only and noise-only audio spectra to generate a gain mask, optimizing noise suppression and signal quality through computation-friendly filters.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If traditional voice activity detection and power spectral density estimation methods are used for audio enhancement, then implementation is simpler, but they fail to efficiently extrapolate time-varying audio data and require expensive expert fine-tuning
Solution Approach 1:
The patent transforms the audio enhancement problem from traditional parameter-based estimation to a deep learning approach where the model learns optimal parameters automatically. The DNN model takes noisy audio spectra as input and directly outputs enhanced audio spectra, eliminating the need for manual parameter tuning and expert fine-tuning while efficiently handling time-varying characteristics.
2Reliability
If deep neural network models are trained with complex loss functions to achieve optimal noise suppression, then noise suppression performance improves, but training becomes computationally expensive and slow
Solution Approach 1:
The patent performs preliminary actions by pre-training the DNN model offline with a comprehensive loss function that balances noise suppression and signal quality. The pre-trained model captures optimal noise suppression strategies during the offline phase, enabling fast real-time inference without requiring complex computational resources during deployment. This separates the computationally intensive training phase from the efficient inference phase.
Solution Approach 2:
The patent implements a dynamic loss function that adaptively balances noise suppression and signal quality preservation during training. The loss function dynamically adjusts weights based on the input audio characteristics, allowing the model to learn optimal suppression strategies for different noise conditions. This dynamic approach improves training efficiency by focusing computational resources on the most critical learning scenarios.
3Object-affected harmful factors
If aggressive noise suppression is applied to remove more noise, then noise removal effectiveness improves, but signal distortion increases
Solution Approach 1:
The patent uses a multi-component loss function that changes the optimization parameters to simultaneously minimize noise while preserving signal quality. The loss function includes multiple terms that balance noise suppression and signal fidelity, allowing the model to learn the optimal trade-off point. This parameter-based balancing enables the model to achieve effective noise removal without excessive signal distortion by learning from diverse training examples with varying noise levels and types.
Data Source
AI summary
A system includes memory storing instructions and a processing device coupled to the memory. The processing device executes the instructions to: receive a noisy audio signal from an audio receiver; pass the noisy audio signal through a deep neural network (DNN) model to generate a mask of a magnitude spectrogram of the noisy audio signal; retrieve a clean-only audio spectra and a noise-only spectra of one or more frequencies that exist within the noisy audio signal; generate a loss function as a combination of the clean-only audio spectra multiplied by the mask and the noise-only spectra multiplied by the mask; and train the DNN model while minimizing the loss function to generate a trained DNN model useable in audio noise suppression and dereverberation.