Neural Audio Encoding With Timing-Preserving Feature Quantization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing audio encoding and decoding technologies result in large volume encoded data, which is inefficient for transmission and storage, and often lead to significant loss of original audio information due to insufficient consideration of timing information.
Innovation Solution
Utilize a combination of convolutional neural networks and recurrent neural networks in both encoding and decoding processes to extract detailed features and timing information, followed by feature quantization, to compress and transmit audio data effectively.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If traditional preset algorithms are used to encode audio data, then the encoding process is simple, but the encoded data volume is large and transmission efficiency is poor
Solution Approach 1:
The patent transforms audio data from time-domain parameters to frequency-domain parameters using Fourier transform, and then applies Mel-scale filtering to map frequency parameters to perceptual parameters. This parameter transformation enables more efficient compression by representing audio data in a form that better matches human perception and reduces redundancy.
Solution Approach 2:
The patent extracts only the essential perceptual features from the audio signal through Mel-filterbank processing and vector quantization. Instead of encoding the complete audio waveform, only the most important spectral parameters are extracted and transmitted, significantly reducing data volume while preserving perceptual quality.
2Quantity of substance
If traditional compression methods are used to reduce encoded data volume, then data volume decreases, but significant information loss occurs in the reconstructed audio
Solution Approach 1:
The patent applies perceptual parameter transformation through Mel-scale filtering and vector quantization to represent audio data in a compressed parameter space. This transformation maintains essential perceptual information while reducing data volume, as the parameters are specifically chosen to preserve human-audible characteristics.
Solution Approach 2:
The patent employs an iterative optimization process during training where the encoder and decoder are jointly trained to minimize reconstruction error. The feedback loop allows the system to learn optimal parameter representations that balance compression efficiency with information preservation, ensuring that decoded audio closely matches the original.
3Loss of information
If conventional encoding methods are used, then the encoding process is fast and simple, but timing information is insufficiently preserved
Solution Approach 1:
The patent performs preprocessing operations including Fourier transform and Mel-filterbank analysis before the main encoding process. These preliminary transformations organize the audio data in a structured frequency-time representation that makes timing information more accessible and easier to preserve during subsequent compression steps.
Data Source
AI summary
This disclosure relates to an audio encoding and decoding method and apparatus, and an electronic device. The audio encoding method includes: acquiring a frame of audio data to be processed; performing an encoding processing on the audio data based on a plurality of encoder blocks; wherein each encoder block of the plurality of encoder blocks comprises a first convolutional neural network and a recurrent neural network; and performing a feature quantization processing on a result after the encoding processing to obtain target encoded data of the audio data.