Normalizing Flow Audio Coding With Split-Band Bit Rate Control
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing audio compression techniques using deep learning networks, such as WaveGlow, face performance limitations in generating high-quality voice waveforms, and there is a need for improved methods that provide excellent audio quality and efficient compression.
Innovation Solution
A method and apparatus using a normalizing flow-based generative model with N flow blocks and (N−1) split blocks for encoding, and N inverse-flow blocks and (N−1) gather blocks for decoding, along with quantization and inverse-quantization blocks, to enhance audio encoding and decoding processes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If WaveGlow uses a flow-based generative model with Mel-spectrogram as condition vector, then voice waveform generation is efficient, but performance is limited in generating high-quality original voice
Solution Approach 1:
The patent divides the normalizing flow into multiple flow blocks (N flow blocks) that process different frequency bands separately. Each flow block handles a specific segment of the frequency spectrum, allowing independent optimization of compression and quality for each segment, thereby resolving the contradiction between generation efficiency and audio quality.
Solution Approach 2:
The patent introduces a split block that divides the transformation signal into multiple frequency bands (first split signal and second split signal), adding a frequency dimension to the processing. This allows the system to handle different frequency components with different precision and compression ratios, improving overall audio quality while maintaining efficiency.
2Manufacturing precision
If traditional audio codecs use psychoacoustic models and bit allocation, then compression restoration quality is indistinguishable from original at 7-10 times compression rate, but the method is complex and computationally intensive
Solution Approach 1:
The patent replaces traditional mechanical signal processing methods (MDCT, psychoacoustic models, bit allocation) with a deep learning-based normalizing flow model. The generative model directly learns the compression and reconstruction mappings from data, eliminating the need for complex hand-crafted algorithms while achieving comparable or superior quality.
Solution Approach 2:
The patent changes the fundamental parameters of audio processing by using a learnable transformation model instead of fixed psychoacoustic models. The normalizing flow adapts its parameters during training to optimize both compression ratio and quality, providing a flexible alternative to rigid traditional algorithms.
3Productivity
If deep learning networks replace traditional signal processing, then compression efficiency is improved, but training complexity and computational resources increase
Solution Approach 1:
The patent segments the deep learning model into multiple flow blocks that can be trained independently and processed in parallel. This segmentation allows the model to achieve high compression efficiency while distributing the computational burden across multiple smaller units, reducing the energy required for training compared to a single monolithic network.
Data Source
AI summary
A method for encoding an input signal using N flow blocks (N is a natural number greater than or equal to 2) and (N−1) split block(s), which is performed by a processor, may comprise: transmitting, by a k-th flow block (k is a natural number greater than or equal to 1 and less than or equal to N−1) among the N flow blocks, a k-th transformation signal obtained by transforming a received signal into a latent representation to a k-th split block among the (N−1) split block(s); splitting, by the k-th split block, the k-th transformation signal by a predetermined ratio, into a first split signal and a second split signal; transmitting, by the k-th split block, the first split signal to a (k+1)-th flow block; and quantizing a signal transformed by an N-th flow block and the second split signals using a quantization block.