Long-Term Prediction State Management in Speech Coding

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing speech coding technologies face inefficiencies in long-term prediction due to fluctuations in LPC coefficients, leading to increased energy in LTP residuals and error propagation issues during packet loss.

Innovation Solution

The method involves modifying the LTP state by re-whitening it using updated LPC coefficients, and scaling down the LTP state to reduce energy and mitigate error propagation, thereby improving coding efficiency and resilience to packet loss.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If LTP prediction is used to exploit periodicity in speech signals, then coding efficiency is improved, but error propagation occurs during packet loss

Engineering Contradiction:
Improvecoding efficiencyVSAvoidresistance to packet loss errors
Core Design Contradiction:
ProductivityVSReliability

Solution Approach 1:

The LTP state is segmented into multiple vectors that are processed independently. During packet loss, only the affected segment needs to be concealed while other segments remain valid, limiting error propagation. The segmentation allows selective updating of LTP state vectors based on packet reception status.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The LTP state vectors are pre-computed and stored before transmission. At the receiver, these pre-computed vectors can be directly used for prediction without real-time computation, enabling faster error concealment and reducing the impact of packet loss on overall system performance.

Inventive Principle:
Principle #10Preliminary action

2Measurement precision

If LPC coefficients are updated frequently to improve spectral accuracy, then measurement precision is improved, but energy in LTP residuals increases

Engineering Contradiction:
Improvespectral accuracyVSAvoidenergy in LTP residuals
Core Design Contradiction:
Measurement precisionVSLoss of energy

Solution Approach 1:

Different parts of the LTP state are treated with different update frequencies. The LPC coefficients used for LTP state computation are updated at specific rates optimized for each local requirement, rather than uniformly across all parameters. This allows spectral accuracy where needed while minimizing residual energy elsewhere.

Inventive Principle:
Principle #3Local quality

Solution Approach 2:

The patent changes the parameters of LTP state vectors by applying transformations such as whitening and scaling. These parameter changes optimize the energy distribution in the residual signal while maintaining spectral accuracy, effectively resolving the contradiction between precise spectral representation and minimal residual energy.

Inventive Principle:
Principle #35Parameter changes

3Productivity

If LTP state is transformed (whitening and scaling) to reduce residual energy, then coding efficiency is improved, but device complexity increases

Engineering Contradiction:
Improvecoding efficiencyVSAvoidencoding complexity
Core Design Contradiction:
ProductivityVSDevice complexity

Solution Approach 1:

The complex whitening and scaling transformations are extracted as separate, modular processing stages. Each transformation is implemented as an independent function that can be optimized separately, reducing overall system complexity while maintaining the benefits of reduced residual energy.

Inventive Principle:
Principle #2Taking out (Extraction)

Data Source

PatentEP2384508B1Speech coding
Publication Date: 2018.09.05 SKYPE
  • EP2384508B1 patent drawingFigure 1a~1b
  • EP2384508B1 patent drawingFigure 2a~4e
  • EP2384508B1 patent drawingFigure 3a

AI summary

A method, system and computer program for encoding speech according to a source-filter model. The method comprises deriving a spectral envelope signal representative of a modelled filter and a first remaining signal representative of a modelled source signal, and deriving a second remaining signal from the first remaining signal by, at intervals during the encoding: exploiting a correlation between approximately periodic portions in the first remaining signal to generate a predicted version of a later portion from a stored version of an earlier portion, and using the predicted version of the later portion to remove an effect of said periodicity from the first remaining signal. The method further comprises, once every number of intervals, transforming the stored version of the earlier portion of the first remaining signal prior to generating the predicted version of the respective later portion.