Neural Network Residual Coding for Low Bit-Rate Voice
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing voice codecs fail to simultaneously achieve high quality, low bit-rate, low complexity, and low memory requirements, with conventional linear predictive-based codecs either providing high quality but not low bit-rate or complexity/memory, or vice versa.
Innovation Solution
Replacing the fixed codebook in a conventional voice decoder with an artificial intelligence model, such as a neural network, to model the residual signal and excite long-term and short-term prediction filters, reducing the bit requirement for residual coding.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a fixed codebook is used in conventional voice codecs for residual coding, then speech quality can be maintained, but bit-rate increases and computational complexity increases
Solution Approach 1:
The patent replaces the traditional fixed codebook mechanism with a neural network model. The neural network learns to directly predict residual signal samples from spectral features, eliminating the need for exhaustive codebook searches and fixed codebook structures. This substitution transforms the mechanical lookup and selection process into an intelligent prediction process, achieving lower bit-rate (only spectral features need to be transmitted) while maintaining speech quality through the network's learned prediction capability.
Solution Approach 2:
The invention changes the fundamental parameters of residual coding by transitioning from representing residual signals through fixed codebook indices to representing them through neural network predictions based on spectral features. This parameter transformation allows the system to encode residual information more efficiently by leveraging the network's ability to capture complex patterns in fewer bits, directly addressing the bit-rate quality tradeoff.
2Measurement precision
If a fixed codebook is used in conventional voice codecs for residual coding, then speech quality can be maintained, but computational complexity increases
Solution Approach 1:
The patent replaces the computationally intensive fixed codebook search mechanism with a neural network prediction system. During decoding, instead of performing exhaustive searches through the fixed codebook to find the best matching residual, the system uses the neural network to directly predict residual samples from spectral features. This substitution dramatically reduces computational complexity while preserving speech quality, as the network performs parallel computations that are more efficient than sequential codebook searches.
3Measurement precision
If a fixed codebook is used in conventional voice codecs for residual coding, then speech quality can be maintained, but memory requirements increase
Solution Approach 1:
The patent extracts and removes the fixed codebook structure from the decoding system. By eliminating the need to store large fixed codebooks in memory, the invention significantly reduces memory requirements. The neural network model processes spectral features directly to generate residual predictions without requiring the decoder to maintain extensive codebook tables, thereby achieving low memory operation while preserving speech quality through the network's learned representations.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
Techniques are described for coding audio signals. For example, using a neural network, a residual signal is generated for a sample of an audio signal based on inputs to the neural network. The residual signal is configured to excite a long-term prediction filter and/or a short-term prediction filter. Using the long-term prediction filter and/or the short-term prediction filter, a sample of a reconstructed audio signal is determined. The sample of the reconstructed audio signal is determined based on the residual signal generated using the neural network for the sample of the audio signal.