MindVoice: Reconstructing Intelligible Speech from Non-invasive Neural Signals with Pretrained Priors
Researchers have developed MindVoice, a novel framework for reconstructing intelligible speech from non-invasive neural signals. This system utilizes pretrained models to overcome the limitations of noisy and blurred neural recordings, which have previously resulted in unintelligible speech synthesis. MindVoice disentangles the reconstruction process into semantic content recovery and acoustic attribute estimation, which are then combined with advanced speech generation and voice cloning techniques. Experiments using EEG and MEG data show significant improvements over existing methods, paving the way for advancements in auditory neuroscience and non-invasive brain-computer interfaces. AI
IMPACT Enables more natural and intelligible communication for brain-computer interfaces and advances auditory neuroscience research.