Audio Neural Network Processing With Sinusoidal Decoder Activation
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 while managing high computational load and complexity.
Innovation Solution
A deep-learning-based generator with an encoder and decoder stage, utilizing sinusoidal activation functions in the decoder layers to model periodic functions, and a training mechanism in a Generative Adversarial Network (GAN) setting to enhance audio processing efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If deep convolutional GANs are used to reduce coding artifacts and noise, then audio processing quality is improved, but computational load and model complexity increase significantly
Solution Approach 1:
The model is divided into an encoder stage and a decoder stage with distinct functions. The encoder extracts latent features from the noisy audio signal, while the decoder reconstructs the clean signal. This segmentation allows each stage to be optimized independently, reducing overall complexity while maintaining processing quality.
Solution Approach 2:
The patent transforms the audio processing problem into a different dimensional space by mapping the audio signal to a latent feature space representation. This dimensional transformation allows the model to capture essential audio characteristics with fewer parameters, reducing computational load while preserving processing quality.
2Manufacturing precision
If deep convolutional GANs are used to reduce coding artifacts and noise, then audio processing quality is improved, but computational load increases significantly
Solution Approach 1:
The patent extracts only the essential features needed for audio reconstruction by mapping the input signal to a compressed latent feature space. This extraction process removes redundant information, allowing the model to achieve high processing quality with reduced computational requirements compared to processing the full-dimensional signal.
Solution Approach 2:
The use of sinusoidal activation functions introduces periodic bias with specific parameter characteristics that are well-suited for modeling tonal content. This parameter choice improves efficiency by requiring fewer computational resources to achieve the same processing quality compared to generic activation functions.
3Device complexity
If standard activation functions are used in the decoder, then model simplicity is maintained, but periodic functions in tonal content cannot be modeled effectively
Solution Approach 1:
The patent changes the activation function parameters to use sinusoidal functions with specific periodic characteristics. This parameter change enables the model to effectively represent tonal content while maintaining relative simplicity in the overall architecture, achieving a balance between model complexity and processing quality.
Solution Approach 2:
Sinusoidal activation functions inherently provide periodic action that matches the periodic nature of tonal audio content. This periodicity allows the model to capture speech and music characteristics effectively without requiring complex architectural modifications, maintaining simplicity while improving tonal content processing.
Data Source
Figure 1
Figure 2a~2b
Figure 3a~3b
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.