Machine-Learning AGC for Stable Speech Gain in Noisy Channels
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Modern automatic gain control systems in conferencing technologies often fail to distinguish between weak speech and noise, leading to undesirable gain changes that can result in unintelligible or incomprehensible audio outputs, as they rely on binary noise versus speech decisions and preset output signal levels rather than considering the desired signal level or room conditions.
Innovation Solution
The implementation of a machine learning-based method that uses a trained model to differentiate between speech and noise, removing background noise to estimate the desired signal level and adjust the gain accordingly, thereby enhancing audio quality and stability in conferencing systems.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional automatic gain control systems use binary noise versus speech decisions and preset output signal levels, then the system structure remains simple, but the audio output quality deteriorates due to undesirable gain changes and inability to distinguish between weak speech and noise
Solution Approach 1:
The patent replaces traditional mechanical/audio-based gain control mechanisms with a machine learning model that processes audio signals. The model uses spectral analysis and classification algorithms to distinguish speech from noise, substituting the simple binary decision mechanism with an intelligent system that considers multiple acoustic features, thereby improving audio output quality while accepting increased system complexity.
Solution Approach 2:
The patent changes the parameters used for gain control from simple preset signal levels to dynamic estimates derived from machine learning model outputs. The system continuously adjusts gain based on the model's speech probability estimates and spectral characteristics, allowing adaptive parameter changes that improve reliability in distinguishing weak speech from noise.
2Reliability
If traditional AGC systems apply gain adjustments based on preset levels without considering desired signal level, then the system operation remains straightforward, but the audio intelligibility deteriorates due to gain changes on weak speech signals
Solution Approach 1:
The patent implements a feedback mechanism where the machine learning model continuously analyzes the audio signal and provides speech probability estimates that feed back into the gain control decision. This closed-loop system adjusts gain based on real-time speech detection, improving audio intelligibility by preventing gain reduction on weak speech signals while maintaining straightforward operation through automated model-based decisions.
Solution Approach 2:
The system uses the machine learning model to automatically determine when gain adjustment is appropriate, making the system self-serve its own control decisions. The model independently evaluates spectral features and speech probability, eliminating the need for complex manual operation or preset configuration, thereby improving intelligibility while maintaining ease of operation.
3Reliability
If traditional AGC systems dampen loud background noise, then the overall signal level is normalized, but the signal-to-noise ratio deteriorates when weak speech is present
Solution Approach 1:
The patent segments the audio signal into different frequency components using spectral analysis, allowing the machine learning model to evaluate speech and noise characteristics in different frequency bands independently. This segmentation enables selective gain adjustment that preserves speech components while attenuating noise, improving signal-to-noise ratio without requiring overly complex full-spectrum processing.
Solution Approach 2:
The patent applies different gain adjustments to different frequency components based on local spectral characteristics identified by the machine learning model. Instead of uniform gain control, the system applies local quality adjustments where speech components are preserved or enhanced while noise components are attenuated, improving signal-to-noise ratio with manageable processing complexity.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
Method includes receiving, through a plurality of channels, audio data corresponding to a plurality of frequency ranges; determining, for each channel's frequency ranges, speech audio and/or noise energy level using a model trained by machine learning; determining a speech signal with removed noise for each channel; determining one or more statistical values associated with an energy level of a channel's speech signal with the removed noise; determining a strongest channel that has highest statistical values associated with an energy level of a speech signal; determining that the one or more statistical values associated with the energy level of the strongest channel's speech signal satisfy a threshold condition; comparing statistical values associated with an energy level of a speech signal of each channel with those of the strongest channel; and determining whether to update a gain value for a channel based on the channel's statistical values associated with the energy level.