A time-domain end-to-end network (IFENet) that extracts one speaker's clean speech from a mixture using the listener's EEG as the cue for which speaker to attend to, replacing the reference recording that ordinary target speaker extraction needs. It keeps the encoder / separator / decoder structure of Conv-TasNet but rebuilds both feature encoders, because one-dimensional convolutions have too small a receptive field to carry long-range speech dependence and too little depth to pull speaker identity out of noisy EEG. SpeechBiMamba passes the encoded mixture and its time-flipped copy through stacked Mamba blocks in a dual-path arrangement, so a selective state-space model captures local and global structure in both directions at cost linear in sequence length. EEGKAN stacks two-head self-attention with Kolmogorov-Arnold Network layers, whose learnable edge activations replace the fixed node activations of a multi-layer perceptron. The two embeddings are fused by convolutional multi-layer cross-attention to estimate a mask, and training maximizes scale-invariant signal-to-distortion ratio.
Target speaker extraction is the cocktail-party task of recovering one speaker's clean speech from a mixture, and the neuro-oriented version replaces a reference recording of that speaker with the li…