ML-Assisted Acoustic Echo Cancellation for Reverberant Speech

The patent introduces an intermediary machine learning model that acts as a bridge between the traditional AEC algorithm and the audio signal. The ML model processes the residual echo signal first, generates probability scores indicating echo presence, and then guides the traditional AEC algorithm's behavior. This intermediary layer enables accurate echo detection in reverberant environments without completely replacing the traditional system, thus balancing accuracy improvement with controlled complexity increase.

The patent segments the echo cancellation task into two distinct parts: (1) ML-based echo detection that analyzes residual signals and outputs probability scores, and (2) traditional AEC algorithm that performs actual signal processing based on ML guidance. This segmentation allows each component to specialize in its strength - ML for accurate detection and traditional algorithms for efficient processing - resolving the contradiction between accuracy and complexity.

The patent implements dynamic adaptation of AEC processing intensity based on real-time ML predictions. The system continuously monitors residual echo probability scores and adjusts the aggressiveness of echo cancellation accordingly. When echo is detected with high probability, aggressive modes are activated; when speech is detected, normal modes are used. This dynamic adjustment resolves the contradiction by making the system adaptive rather than static, preventing over-processing while maintaining effective echo suppression when needed.

The patent incorporates feedback loops where the ML model's predictions about echo presence feed back to control the AEC algorithm's behavior. The system uses the residual signal from AEC processing as input to the ML model, which then generates feedback control signals that adjust subsequent AEC processing intensity. This feedback mechanism ensures that aggressive echo cancellation is applied only when echo is actually present, preventing speech quality deterioration from unnecessary over-processing.

The patent changes the parameter space by transforming raw residual echo signals into probability scores through ML processing. Instead of using fixed threshold-based detection that causes false positives, the ML model learns optimal parameter transformations during training, outputting calibrated probability scores that accurately represent echo likelihood. This parameter transformation resolves the contradiction by providing nuanced detection confidence levels that prevent misclassification of speech as echo.

The patent substitutes the mechanical threshold-based detection mechanism with a machine learning-based probabilistic classification system. The traditional mechanical approach of comparing signal magnitudes against fixed thresholds is replaced by an ML model that learns complex patterns and relationships in the data. This substitution enables more accurate discrimination between echo and speech, reducing false detections while maintaining sensitivity to actual echo content.