Audio Quality Metric Learning with Labeled and Relative Ratings
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing audio quality assessment methods, whether rule-based or learning-based, suffer from lack of generalization and require significant amounts of human-annotated data, leading to inefficiencies and poor performance on out-of-sample data.
Innovation Solution
A semi-supervised deep-learning-based method that utilizes both human-annotated and unlabeled audio samples, employing a neural network trained with a combination of loss functions to predict audio quality scores, allowing for improved generalization and reduced reliance on expensive listener data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If rule-based systems are used for audio quality assessment, then the system is perceptually-motivated and interpretable, but the system has narrow focus on specific types of signals or degradations and lacks generalization
Solution Approach 1:
The patent combines rule-based perceptual modeling with machine learning techniques to create a hybrid system. The neural network is trained using perceptually-motivated features and loss functions that incorporate human listening test data, merging the interpretability of rule-based systems with the adaptability of learning-based systems to achieve both goals simultaneously
Solution Approach 2:
The patent develops a universal audio quality assessment framework that can handle multiple types of signals and degradations through a single trained model. The system uses perceptual features that are broadly applicable across different audio types, enabling the model to generalize to out-of-sample data while maintaining perceptual relevance
2Adaptability or versatility
If learning-based systems are used for audio quality assessment, then the system is easy to repurpose to other tasks and degradations, but the system requires considerable amounts of human annotated data
Solution Approach 1:
The patent performs preliminary action by pre-training the neural network on large amounts of perceptual feature data before fine-tuning with human annotated data. This allows the system to learn general audio representations that can be adapted to specific quality assessment tasks with minimal annotated data, reducing the overall annotation requirement
Solution Approach 2:
The patent introduces perceptual features and intermediate representations as mediators between raw audio data and quality predictions. These intermediate features capture essential audio characteristics that transfer across different tasks and degradations, enabling the system to be repurposed for different applications without requiring complete retraining with new annotated data
3Productivity
If traditional objective metrics are used for audio quality assessment, then the system can scale up automatic quality assessment with consistent results, but the system relies on intricate and complex rule-based processing
Solution Approach 1:
The patent replaces complex mechanical rule-based processing systems with a neural network-based system. The neural network automatically learns the complex transformations and relationships between audio features and quality metrics, substituting hand-crafted rule-based systems with a learned model that achieves comparable or better performance with potentially simpler architecture
Data Source
AI summary
Described is a method of training a neural-network-based system for determining an indication of an audio quality of an audio input. The method includes obtaining, as input, at least one training set comprising audio samples. The audio samples include audio samples of a first type and audio samples of a second type, wherein each of the first type of audio samples is labelled with information indicative of a respective predetermined audio quality metric, and wherein each of the second type of audio samples is labelled with information indicative of a respective audio quality metric relative to that of a reference audio sample. The method further includes: inputting the training set to the neural-network-based system; and iteratively training the system to predict the respective label information of the audio samples in the training set.