Multi-source Localization Based on Graph Signal Processing

‍ ‍The Direction of Arrival （DOA） estimation technique is an important tool in speech enhancement and acoustic detection. It has important applications， such as speech robots， video conferencing， hearing aids， and sonar. Recently developed DOA estimation methods， such as Graph Signal Processing （GSP） methods， have demonstrated excellent angle estimation capabilities， offering potential for improved solutions for source DOA estimation. However， in multi-source scenarios， the GSP algorithm fails to directly obtain the orthogonal complement matrix of the received signal feature vectors from the adjacency matrix， rendering it ineffective in such situations. To address this limitation， this paper proposes a multi-source separation based on frequency domain single-source region detection for wideband speech signals， followed by the utilization of GSP and clustering algorithms for wideband multi-source localization. Specifically， this paper first extends the GSP method to the frequency domain； then single-source dominant regions for frequency domain GSP single-source localization are identified by employing a short-time Fourier transform to divide the signal into several time-frequency regions； finally， all localization results are clustered， and the final angle estimation is obtained through weighted averaging. We used the LibriSpeech speech corpus to construct acoustic source signals for multi-source localization simulation. The simulation results demonstrate that our proposed method outperforms other algorithms， with errors being kept within 3° under high signal-to-noise ratio conditions. Additionally， we utilized a circular six-microphone array to conduct localization measurements on several sets of recorded audio data using the proposed algorithm. The results show that the proposed algorithm achieves smaller localization errors and performs better at distinguishing sources when the sources are closer.