Audio Enhancement Gain Prediction for Distorted Frequency Bands
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing audio enhancement methods struggle with handling various encoding/transcoding processes, noise, and reverberation in degraded audio signals, leading to unsatisfactory listening experiences and complexity issues, especially in real-time applications.
Innovation Solution
A supervised training method for a machine learning model (MLM) that calculates gains for frequency bands, using pre-processing to equalize frequency energy distributions and a loss function to prevent over-suppression, reducing computational complexity and enhancing audio quality.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional audio enhancement algorithms are used, then audio quality can be improved to some extent, but the algorithms become overly complex and difficult to implement in real-time
Solution Approach 1:
The patent replaces complex traditional signal processing algorithms with a machine learning-based system that uses pre-trained models to predict and remove artifacts. Instead of using computationally intensive real-time processing, the system substitutes a simpler inference process that leverages pre-computed neural network weights, significantly reducing implementation complexity while maintaining enhancement effectiveness
Solution Approach 2:
The system performs preliminary training of the machine learning model offline using extensive training data and pre-processing computations. By preparing the model weights and processing parameters in advance, the actual real-time execution requires minimal computational resources, effectively separating the complex training phase from the simple inference phase
2Reliability
If the enhancement algorithm processes all frequency bands equally, then comprehensive enhancement is achieved, but computational resources are wasted on already-clean frequency bands
Solution Approach 1:
The patent applies different processing strategies to different frequency bands based on their specific characteristics and degradation patterns. The system identifies which frequency bands contain artifacts and applies enhancement only to those bands, while leaving already-clean bands unchanged. This localized approach optimizes computational energy by focusing resources only where needed rather than uniformly processing all frequency content
3Adaptability or versatility
If the system handles multiple codec chains, then versatility is improved, but the processing time and complexity increase significantly
Solution Approach 1:
The system employs a universal machine learning model that has been trained to recognize and handle artifacts from multiple different codec chains simultaneously. Rather than implementing separate processing pipelines for each codec, the single model learns to identify and correct degradation patterns characteristic of various codecs, providing multi-functional capability across different encoding scenarios without requiring codec-specific processing logic
Solution Approach 2:
The system adapts to different codec chains by detecting their specific artifact characteristics and adjusting the enhancement parameters accordingly. The machine learning model dynamically modifies its processing behavior based on the detected degradation pattern, allowing it to optimize performance for each specific codec while maintaining a unified processing architecture that avoids the time cost of multiple specialized pipelines
Data Source
AI summary
The present disclosure relates to the field of audio enhancement, and in particular to methods, devices and software for supervised training of a machine learning model, MLM, the MLM trained to enhance a degraded audio signal by calculating gains to be applied to frequency bands of the degraded audio signal. The present disclosure further relates to methods, devices and software for use of such a trained MLM.