Audio Recognition False Positive Reduction via Sound Class Comparison

The patent replaces the mechanical signal processing approach (subtractor circuitry operating on audio signals) with an information processing approach (machine learning model operating on sound class labels). Instead of manipulating audio waveforms directly, the system uses trained models to classify sounds and compare class labels, eliminating the need for complex parameter adjustments related to gain, phase, and echo cancellation.

The patent changes the operating parameters from continuous audio signal values requiring precise gain and phase adjustment to discrete sound class labels that can be directly compared. The machine learning models output probability distributions over sound classes, which are then compared at the class level rather than requiring pixel-perfect signal subtraction.

The patent applies preliminary action by training machine learning models beforehand to recognize and classify different sound types. Instead of attempting to separate sounds in real-time through complex signal processing, the system pre-trains models to identify sound characteristics, then uses these pre-trained models to classify sounds and compare class labels, which is much more reliable than real-time signal subtraction.

The patent introduces sound class labels as an intermediary between the raw audio signals and the final discrimination decision. Rather than directly comparing and subtracting audio signals which are mixed in unpredictable ways, the system uses machine learning models to translate signals into class labels, then compares these labels to determine whether a sound is artificial or real-life.