Self-Attention Audio Denoising for Multi-Channel Spectrum

The patent replaces the recurrent neural network (RNN) mechanism with a self-attention mechanism. The self-attention mechanism computes mixing coefficients through parallel attention computations across all time-frequency bins simultaneously, eliminating the sequential processing bottleneck of RNNs while maintaining or improving denoising performance through better spectral representation.

The patent transforms the processing approach by introducing a global attention dimension that considers all time-frequency bins across all channels simultaneously. The self-attention mechanism computes attention weights that capture long-range dependencies and global spectral patterns, adding a dimensional perspective that RNNs cannot achieve through their sequential nature.

The self-attention mechanism implements a feedback loop where each time-frequency bin's representation is continuously refined by attending to all other bins. The attention weights are computed based on queries, keys, and values from all channels, and the resulting attended representations are fed back to update the mixing coefficients, creating an iterative refinement process that preserves spectral information.

The self-attention mechanism serves multiple functions simultaneously: it estimates mixing coefficients, represents multi-channel spectrum information, captures temporal and spectral dependencies, and performs denoising. This universal approach replaces the specialized RNN architecture with a single mechanism that handles all these tasks through its attention-based computation.