Hearing Device Steering AI Training with Mixed Speech-Noise Data

The patent applies preliminary action by generating and using synthetic acoustic datasets before deploying the hearing device in real-world noisy environments. The system pre-trains machine learning models using mixed speech and noise recordings created in controlled conditions, allowing the algorithm to learn optimal steering behavior in advance. This pre-training approach enables the device to handle real noisy scenarios more effectively without requiring extensive real-world data collection.

The patent employs copying by creating synthetic representations of real-world acoustic environments. Instead of relying solely on actual recorded noisy scenarios, the system generates artificial acoustic datasets that replicate the statistical and spectral characteristics of real noisy environments. These synthetic copies are used to train the machine learning models, providing a scalable and controlled training approach that mimics diverse real-world conditions.

The patent replaces mechanical data collection methods with computational synthesis. Instead of physically recording diverse acoustic scenarios across multiple environments and conditions, the system uses computational algorithms to generate synthetic acoustic datasets. This substitution of mechanical recording processes with computational generation reduces the logistical complexity of data collection while maintaining dataset diversity through algorithmic variation of acoustic parameters.

The patent applies parameter changes by systematically varying acoustic parameters in synthetic dataset generation. The system modifies parameters such as noise levels, speech directions, reverberation characteristics, and signal-to-noise ratios to create diverse training scenarios. This parametric approach allows comprehensive coverage of acoustic variations without requiring proportional increases in physical data collection resources, managing complexity through controlled parameter manipulation.