Direction of Arrival Estimation Using Acoustic Intensity Vectors

The patent segments the DOA estimation process into multiple interpretable stages: acoustic intensity vector calculation from microphone signals, reverberation component estimation using DNN, noise suppression mask generation, and final DOA derivation. Each stage produces intermediate results that can be analyzed and understood, maintaining transparency while achieving robustness through the structured multi-stage approach.

The patent introduces intermediate representations (acoustic intensity vectors, reverberation components, noise suppression masks) that serve as mediators between the raw input signals and the final DOA estimation. These intermediates make the learning process transparent by providing interpretable stages, while the DNN-based reverberation estimation maintains robustness against SNR variations.

The patent extracts and removes the reverberation component from the acoustic intensity vector using DNN-based estimation. By separating the reverberation component from the direct sound, the system improves DOA estimation accuracy while managing computational complexity through targeted processing of specific signal components rather than entire signal streams.

The patent applies noise suppression masks selectively in the time-frequency domain rather than uniformly across all frequencies and time frames. This partial action approach improves accuracy for relevant frequency regions while reducing unnecessary computational operations in regions where noise suppression is less critical, thereby balancing accuracy improvement with computational complexity management.