Neural Network Audio Signal Processing with Psychoacoustic Loss
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing audio signal processing methods using neural network models often result in signal loss during encoding and decoding, necessitating a technique to minimize this loss and improve the quality of restored audio signals by considering psychoacoustic factors.
Innovation Solution
The method involves training neural network models using a psychoacoustic model to calculate differences between initial and final audio signals in both time and frequency domains, updating parameters to minimize these differences, and applying masking thresholds to reduce noise, utilizing multiple loss functions to optimize the encoding and decoding process.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If traditional encoding and decoding methods are used, then processing speed is maintained, but signal loss occurs and audio quality deteriorates
Solution Approach 1:
The patent changes the training parameters and loss function parameters of neural network models by incorporating psychoacoustic parameters (masking thresholds, critical bands) to optimize the encoding-decoding process. This allows the system to learn transformations that preserve perceptually important signal characteristics while discarding imperceptible information, thereby reducing signal loss without compromising quality.
Solution Approach 2:
The patent introduces psychoacoustic models as an intermediary layer between the encoding and decoding processes. These models provide masking thresholds and critical band information that guide the neural networks in making compression decisions, acting as a mediator that bridges signal processing and perceptual optimization.
2Manufacturing precision
If neural network models are trained without psychoacoustic factors, then training complexity is reduced, but noise is generated and quality of restored audio signal deteriorates
Solution Approach 1:
The patent makes the loss function multi-functional by combining multiple objectives: minimizing time-domain error, minimizing frequency-domain error, and minimizing perceptually-weighted error based on psychoacoustic parameters. This universal loss function guides the neural network to simultaneously achieve multiple goals, improving restored signal quality while the modular structure keeps training manageable.
Solution Approach 2:
The patent performs preliminary computation of psychoacoustic parameters (masking thresholds, power spectral density, critical bands) before the actual training process. These pre-computed parameters are then used to weight the loss function during training, allowing the network to focus on perceptually important regions without adding real-time computational complexity during inference.
3Manufacturing precision
If multiple loss functions are used to train neural network models, then audio signal quality is improved, but computational time and processing complexity increase
Solution Approach 1:
The patent segments the loss function into distinct components: time-domain loss, frequency-domain loss, and psychoacoustic-weighted loss. Each component can be computed and optimized separately, allowing for efficient training through staged optimization or selective emphasis on different components based on specific application requirements.
Solution Approach 2:
The patent implements continuous optimization by combining multiple loss functions into a unified training objective that operates throughout the entire training process. Rather than sequential processing, the multi-component loss function provides continuous guidance to the neural network, maintaining optimization pressure across all dimensions (time, frequency, and perceptual) simultaneously.
Data Source
AI summary
A method and apparatus for processing an audio signal are disclosed. According to an example embodiment, a method of processing an audio signal may include acquiring a final audio signal for an initial audio signal using a plurality of neural network models generating output audio signals by encoding and decoding input audio signals, calculating a difference between the initial audio signal and the final audio signal in a time domain, converting the initial audio signal and the final audio signal into Mel-spectra, calculating a difference between the Mel-spectra of the initial audio signal and the final audio signal in a frequency domain, training the plurality of neural network models based on results calculated in the time domain and the frequency domain, and generating a new final audio signal distinguished from the final audio signal from the initial audio signal using the trained neural network models.