Multi-Scale Codebook Search for Low-Complexity Audio Coding
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current speech and audio coding technologies face challenges in reducing memory and computational complexity, particularly in the nearest neighbor search within multiple scale lattice codebooks, which limits their efficiency in mobile terminal communications.
Innovation Solution
The method involves determining potential basis code vectors for each set of basis code vectors associated with scale representatives, calculating distortion metrics based on these vectors, and selecting code vectors for encoding input vectors, thereby reducing search complexity and memory requirements.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If a multiple scale lattice codebook with several lattice truncations is used for speech and audio coding, then coding efficiency is improved, but memory requirements and computational complexity increase
Solution Approach 1:
The codebook search space is segmented into multiple scale levels, where each scale represents a different granularity of code vectors. The search process is divided into coarse-to-fine stages, evaluating distortion metrics at each scale level to progressively narrow down candidate code vectors, thereby reducing the overall computational complexity while maintaining coding efficiency.
Solution Approach 2:
The distortion metric is evaluated preliminarily at each scale level before final code vector selection. By computing distortion metrics at coarser scales first and using these preliminary results to guide the search at finer scales, the method avoids exhaustive search at the finest scale, thus reducing computational complexity while preserving coding accuracy.
2Manufacturing precision
If a multiple scale lattice codebook with several lattice truncations is used for speech and audio coding, then coding efficiency is improved, but memory requirements increase
Solution Approach 1:
The codebook is segmented into multiple scale levels with different numbers of code vectors. By organizing code vectors hierarchically across scales, the system can store fewer total code vectors compared to a single exhaustive codebook, as coarser scales provide approximate representations that reduce the need for storing every possible fine-grained code vector.
Solution Approach 2:
Distortion evaluation is performed preliminarily at each scale level using stored scale representatives and code vector indices. This preliminary evaluation allows the system to identify promising candidates without retrieving and processing all code vectors from the complete codebook, thereby reducing memory access requirements while maintaining coding efficiency.
3Measurement precision
If distortion is evaluated for all leader classes in each lattice truncation at every scale, then coding accuracy is maintained, but computational complexity increases
Solution Approach 1:
The evaluation process is segmented across multiple scale levels, where distortion metrics are computed at each scale rather than evaluating all leader classes simultaneously at the finest scale. This hierarchical segmentation reduces the number of operations at each stage while preserving the ability to achieve accurate coding through progressive refinement.
Solution Approach 2:
Distortion metrics are evaluated preliminarily at coarser scales before final code vector selection at finer scales. This preliminary evaluation at multiple scales allows the system to identify and eliminate poor candidates early, maintaining coding accuracy by ensuring thorough evaluation of promising candidates while avoiding unnecessary computation for clearly suboptimal options.
Data Source
Figure 1a~1b
Figure 2a~2c
Figure 3~5
AI summary
It is inter alia disclosed to determine, for each set of basis code vectors of a plurality of sets of basis code vectors, a potential basis code vector for encoding an input vector, wherein each set of basis code vectors is associated with at least one scale representative of a plurality of scale representatives, and to determine a code vector for encoding the input vector from a subset of code vectors, said subset of code vectors comprising, for each determined potential basis code vector and each scale representative associated with the set of basis code vectors of the respective potential basis code vector, a code vector based on the respective potential basis code vector scaled by the respective scale representative.