Generative Waveform Decoding for Low-Bitrate Audio Reconstruction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing audio coding technologies face challenges in reconstructing perceptually relevant features due to coding artifacts, particularly at low bit-rates, and lack the generative power of models to optimize perceptual performance.
Innovation Solution
A method and system using a generative model that implements a probability density function to decode a finite bitrate representation of a source signal, allowing for seamless integration of waveform and parametric coding, enhancing perceptual performance by filling spectral holes and improving sample distortion measures.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of energy
If low-rate quantization of waveforms is used, then bitrate is reduced, but coding artifacts are introduced that degrade perceptual quality
Solution Approach 1:
The patent introduces a generative model as an intermediary component between the quantizer and the final reconstructed signal. This generative model takes the quantized waveform approximation and transforms it into a more perceptually accurate reconstruction by learning the underlying probability distribution of audio signals, thereby mediating the trade-off between bitrate reduction and perceptual quality preservation
Solution Approach 2:
The patent changes the parameter representation from direct waveform samples to parameters of a probability distribution (mean and standard deviation). By quantizing and transmitting these distribution parameters rather than raw waveform values, the system achieves efficient bitrate reduction while the generative model uses these parameters to synthesize high-quality perceptual reconstructions
2Device complexity
If simple quantizers are used in transform coding, then device complexity is reduced, but manufacturing precision of signal reconstruction is degraded
Solution Approach 1:
The patent replaces traditional mechanical/mathematical quantization systems with a data-driven generative model. Instead of using complex vector quantizers or sophisticated transform coding schemes, the system uses simple quantizers to encode distribution parameters, which are then decoded by a neural network-based generative model trained to produce high-fidelity reconstructions, substituting computational learning for complex quantization algorithms
3Reliability
If bandwidth extension is applied to hide coding artifacts, then perceptual performance is improved, but device complexity increases
Solution Approach 1:
The patent merges the bandwidth extension functionality directly into the generative model. Rather than applying bandwidth extension as a separate post-processing step, the generative model is trained to inherently reconstruct signals with proper spectral characteristics, combining the signal reconstruction and bandwidth extension tasks into a single unified operation that reduces overall system complexity
Data Source
Figure 1~2
Figure 3~4
Figure 5
AI summary
Described herein is a method of waveform decoding, the method including the steps of: (a) receiving, by a waveform decoder, a bitstream including a finite bitrate representation of a source signal; (b) waveform decoding the finite bitrate representation of the source signal to obtain a waveform approximation of the source signal; (c) providing the waveform approximation of the source signal to a generative model that implements a probability density function, to obtain a probability distribution for a reconstructed signal of the source signal; and (d) generating the reconstructed signal of the source signal based on the probability distribution. Described are further a method and system for waveform coding and a method of training a generative model.