Neural Audio Decoding With Sinusoidal Activation for Efficient Restoration
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing deep-learning-based approaches for processing audio signals with tonal content, such as speech and music, face challenges in efficiently removing coding artifacts and noise due to their complexity in terms of model parameters and computational load, and struggle to model periodic functions naturally present in these signals.
Innovation Solution
A deep-learning-based generator with an encoder and decoder stage, utilizing sinusoidal activation in at least one layer of the decoder stage, applies time and channel recurrence to the latent feature space representation, and is trained in a Generative Adversarial Network (GAN) setting to enhance audio signal processing efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If deep convolutional GANs are used for speech denoising and coding artifact reduction, then the ability to process audio signals is improved, but the computational load and model complexity increase significantly
Solution Approach 1:
The patent changes the activation function parameter from conventional ReLU to sinusoidal activation functions (such as Sine, Sinusoidal Linear Unit). This parameter change in the neural network architecture enables the model to inherently represent periodic functions with fewer parameters, thereby reducing model complexity while maintaining or improving audio processing capability for tonal signals like speech and music.
Solution Approach 2:
The patent applies periodic sinusoidal activation functions within the neural network layers to model the periodic nature of tonal audio signals. By using activation functions that naturally represent periodic behavior, the network requires fewer layers and parameters to capture temporal patterns in speech and music, thus reducing overall model complexity while improving reliability for these specific signal types.
2Reliability
If deep convolutional GANs are used for speech denoising and coding artifact reduction, then the ability to process audio signals is improved, but the computational load increases significantly
Solution Approach 1:
By changing the activation function parameter to sinusoidal forms, the network achieves better representation of audio signals with fewer computational operations. The sinusoidal activation functions provide a more efficient parameterization for periodic signals, reducing the number of layers and parameters needed, thereby lowering computational load while maintaining processing capability.
Solution Approach 2:
The use of periodic sinusoidal activation functions aligns with the natural periodicity of audio signals, allowing the network to process tonal content more efficiently. This periodic approach reduces the computational burden by requiring fewer operations to capture the essential characteristics of speech and music signals compared to conventional activation functions.
3Device complexity
If conventional activation functions are used in neural networks, then the model structure is simpler, but the ability to model periodic functions in audio signals is reduced
Solution Approach 1:
The patent directly addresses this contradiction by implementing periodic sinusoidal activation functions (Sine, Sinusoidal Linear Unit) in the neural network layers. These activation functions inherently capture periodic patterns in audio signals, significantly improving the modeling capability for tonal content while maintaining a relatively simple network architecture. The periodic nature of these functions provides a built-in mechanism for representing temporal patterns without requiring complex network structures.
Solution Approach 2:
By changing the activation function parameter from non-periodic (ReLU) to periodic (sinusoidal), the network gains the ability to model periodic functions effectively. This parameter change enables the network to represent temporal patterns in speech and music with fewer parameters and layers, achieving better periodic function modeling capability without substantially increasing model structural complexity.
Data Source
AI summary
Described herein is a method of processing an audio signal using a deep-learning-based generator, wherein the method includes the steps of: (a) inputting the audio signal into the generator for processing the audio signal; (b) mapping a time segment of the audio signal to a latent feature space representation, using an encoder stage of the generator; (c) upsampling the latent feature space representation using a decoder stage of the generator, wherein at least one layer of the decoder stage applies sinusoidal activation; and (d) obtaining, as an output from the decoder stage of the generator, a processed audio signal. Described are further a method for training said generator and respective apparatus, systems and computer program products.