NeuroSonic Reconstructs Speech from EEG Using Conditional Flow Matching.

Researchers have introduced NeuroSonic, a novel framework designed to reconstruct continuous speech directly from electroencephalography (EEG) signals. This method addresses the inherent challenges of EEG, which provides weak and diffuse measurements, by employing a conditional flow-matching approach. Instead of traditional waveform prediction or stochastic denoising, NeuroSonic learns a deterministic probability-flow velocity field. This field effectively transports a noisy acoustic state towards clear speech, guided by EEG conditioning. The system embeds both EEG and audio into a shared token space and processes them with a time-conditioned gated Transformer that parameterizes the transport ordinary differential equation. This architecture allows for explicit modeling of trajectory evolution without requiring iterative stochastic sampling. Evaluations on CineBrain and EAV benchmarks demonstrate NeuroSonic's superior performance. It achieves significant gains in distributional realism, spectral fidelity, and perceptual quality, particularly in segments with high artifact variability, outperforming GAN, diffusion, and mean-flow baselines.