Adaptive Formant Sharpening for Speech Coding

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing speech coding technologies face challenges in maintaining speech quality, especially at low signal-to-noise ratios and in bandwidth extension, due to limitations in formant sharpening techniques and noise modeling in the source-system speech model.

Innovation Solution

The implementation of adaptive formant sharpening techniques, where the formant-sharpening factor is varied based on the long-term signal-to-noise ratio, and the application of specific filters to the fixed codebook vector to emphasize formant regions, thereby improving speech reconstruction quality in both clean and noisy conditions.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Manufacturing precision

If formant sharpening is applied to improve speech reconstruction quality, then speech quality is improved in clean conditions, but artifacts are introduced in noisy conditions and bandwidth extension

Engineering Contradiction:
Improvespeech reconstruction qualityVSAvoidartifacts in noisy conditions
Core Design Contradiction:
Manufacturing precisionVSObject-affected harmful factors

Solution Approach 1:

The patent applies dynamic adaptation of the formant sharpening filter based on noise conditions. The filter is selectively applied or adapted in magnitude and frequency based on the measured noise level, allowing the system to optimize speech reconstruction quality while avoiding artifacts in noisy conditions. This is achieved through noise level detection and conditional filter application.

Inventive Principle:
Principle #15Dynamics

Solution Approach 2:

The patent changes the parameters of the formant sharpening filter based on noise conditions. The filter magnitude and frequency characteristics are adjusted according to the measured noise level, allowing the system to maintain speech quality in clean conditions while reducing artifact introduction in noisy conditions. This parameter adaptation resolves the contradiction between quality improvement and artifact prevention.

Inventive Principle:
Principle #35Parameter changes

2Manufacturing precision

If pitch and formant sharpening are applied to low band excitation for bandwidth extension, then perceptual quality of low-band synthesis is improved, but audible artifacts are more likely to occur in high band synthesis

Engineering Contradiction:
Improveperceptual quality of low-band synthesisVSAvoidaudible artifacts in high band synthesis
Core Design Contradiction:
Manufacturing precisionVSObject-affected harmful factors

Solution Approach 1:

The patent applies dynamic control of formant sharpening in bandwidth extension based on noise level detection. The filter is selectively applied or adapted in magnitude and frequency based on the measured noise level, allowing the system to optimize low-band synthesis quality while avoiding artifact propagation to high band synthesis. This conditional adaptation resolves the contradiction between quality improvement and artifact prevention in bandwidth extension.

Inventive Principle:
Principle #15Dynamics

3Measurement precision

If fixed codebook search is performed to capture aperiodic component, then excitation accuracy is improved, but computational complexity increases

Engineering Contradiction:
Improveexcitation accuracyVSAvoidcomputational complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent applies preliminary pitch sharpening to the excitation signal before performing the fixed codebook search. This pre-processing step enhances the periodic components in the excitation, allowing the fixed codebook search to more effectively capture the remaining aperiodic components. This preliminary action improves excitation accuracy while reducing the computational burden during the actual codebook search by providing a better-initialized signal.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP2951823B1Code-excited linear prediction method and apparatus
Publication Date: 2022.01.26 QUALCOMM INC
  • EP2951823B1 patent drawingFigure 1
  • EP2951823B1 patent drawingFigure 2
  • EP2951823B1 patent drawingFigure 3A~3C

AI summary

A method of processing an audio signal includes determining an average signal-to-noise ratio for the audio signal over time. The method includes, based on the determined average signal-to-noise ratio, a formant-sharpening factor is determined. The method also includes applying a filter that is based on the determined formant-sharpening factor to a codebook vector that is based on information from the audio signal.