Speech Codec Silence Gain Control for Quantization Noise Artifacts
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing speech coding methods fail to effectively reduce noise artifacts in silence areas of speech signals, particularly when a voice activity detector (VAD) is not available, due to large quantization errors exacerbated by fixed-point implementation with integer accuracy limitations.
Innovation Solution
A method that involves obtaining and smoothing the energy level of speech subframes, determining silence areas based on spectral tilt parameters, and modifying the energy levels using a gain factor to reduce noise artifacts, without relying on a VAD, by setting a gain value between 0 and 1 based on the total input level and spectral tilt, and iteratively adjusting the gain for outgoing samples.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If speech compression is applied to reduce bit rate, then bandwidth requirement is reduced, but speech quality degrades
Solution Approach 1:
The patent applies different processing to different parts of the speech signal: silence regions are identified and treated with energy reduction to remove artifacts, while speech regions maintain their original quality. This local differentiation allows low bit-rate compression to preserve perceptual quality by focusing noise reduction only where needed.
Solution Approach 2:
The patent modifies the energy parameter of silence regions by applying a gain factor less than 1.0 to reduce quantization noise artifacts. This parameter change allows the system to maintain low bit-rate operation while improving perceived speech quality by reducing audible artifacts in non-speech portions.
2Device complexity
If fixed-point implementation with integer accuracy is used, then device complexity is reduced, but quantization errors increase in silence areas
Solution Approach 1:
The patent extracts and identifies silence regions from the speech signal using energy-based detection. By separating silence regions from speech regions, the system can apply specialized processing (energy reduction) only to silence areas, effectively removing quantization artifacts without requiring higher precision arithmetic throughout the entire signal processing chain.
Solution Approach 2:
The patent changes the energy parameter of silence regions by applying a gain factor, which reduces the impact of quantization errors in these regions. This parameter modification compensates for the limited precision of fixed-point arithmetic by actively reducing the magnitude of quantization artifacts in silence portions.
3Manufacturing precision
If noise suppression is applied to improve speech quality, then perceptual quality improves, but device complexity increases
Solution Approach 1:
The patent applies noise suppression locally only to identified silence regions rather than to the entire speech signal. By using energy-based detection to distinguish silence from speech, the system applies gain reduction only where artifacts are present, avoiding unnecessary processing in speech regions and thereby reducing overall computational complexity.
Solution Approach 2:
The patent applies partial noise suppression by targeting only silence regions rather than processing the entire signal. This partial action is sufficient to improve perceptual quality since quantization artifacts are primarily audible in silence portions, avoiding the complexity of full-signal processing while achieving the desired quality improvement.
Data Source
AI summary
There is provided a method of reducing effect of noise producing artifacts in silence areas of a speech signal for use by a speech decoding system. The method comprises obtaining a plurality of incoming samples of a speech subframe; summing an absolute value of an energy level for each of the plurality of incoming samples to generate a total input level (gain_in); smoothing the total input level to generate a smoothed level (Level_in_sm); determining that the speech subframe is in a silence area based on the total input level, the smoothed level and a spectral tilt parameter; defining a gain using k1*(Level_in_sm/1024)+(1−k1), where K1 is a function of the spectral tilt parameter; and modifying an energy level of the speech subframe using the gain.