Three-dimensional sound source localization method using hybrid model with three microphones

Aiming at the problem that mobile robot platforms are difficult to deploy large-scale arrays due to limited structural sizes, which results in underdetermined equations and spatial ambiguity in three-dimensional (3D) sound source localization, this paper proposes a hybrid-model stepwise decoupling localization method using a non-collinear three-microphone array with a vertical offset. By creatively leveraging the complementary characteristics of near-field and far-field models, the 3D localization problem is decomposed into three sequential stages. First, a low-resolution global coarse search is conducted using the near-field Steered Response Power with Phase Transform (SRP-PHAT) algorithm to obtain an initial position estimate. Second, a far-field plane-wave approximation model is introduced to decouple the angle and distance parameters. Here, a kurtosis-weighted angular loss function is constructed, and an adaptive-step momentum gradient descent algorithm is used to refine the direction of arrival (DOA). Finally, with the DOA locked, a near-field one-dimensional radial high-resolution search determines the precise distance. Experiments on fixed-bracket and quadruped-robot platforms indicate that the proposed method reduces the influence of environmental reverberation and robot ego-noise under the evaluated conditions. Compared with the tested SRP-PHAT and MUSIC implementations, the method reduces spatial false-peak interference and improves localization stability under the adopted sparse three-microphone configuration.