"Target speech hearing" algorithm at the University of Washington cancels environmental noise, extracts target speech by enrolling speaker with a short, noisy example through binaural microphones.
Enrollment interface requires wearer to look at speaker for neural network training, significantly advancing noise-canceling headphones by selectively picking speakers based on traits.
System optimized TFGridNet to run real-time on embedded CPUs, used synthetic data for generalizable training, solving challenges for speech extraction.
Enrollment step aligns speaker's voice across binaural microphones, creating unique user interface problem solution for hearable application domain.
#target-speech-hearing-algorithm #noise-cancellation #deep-learning #speech-extraction #hearable-technology
Collection
[
|
...
]