Intelligent Hearable System for Target Speech Extraction in Noisy Environments
In crowded settings, the human brain can focus on speech from a target speaker based on prior knowledge of their voice. We introduce a novel intelligent hearable system that replicates this ability, allowing users to hear target speech while ignoring interference. Unlike traditional methods that require clean speech samples for enrollment, our system uses a single, short, noisy binaural recording obtained by the user looking at the target speaker for a few seconds. This noisy example is sufficient for subsequent speech extraction, achieving a 7.01 dB signal quality improvement with less than 5 seconds of audio. Our system processes 8 ms of audio chunks in 6.24 ms on an embedded CPU. User studies show the system generalizes well to various real-world scenarios and does not degrade performance compared to using clean examples. This innovative approach enhances human auditory perception by leveraging artificial intelligence, offering a user-friendly and effective solution for target speech hearing in noisy environments.
