Binaural Multiple Sources Localization with Extracting High Local SNR Frequencies Based on Cepstrum

摘要

This paper investigates the directions of arrival (DOA) estimation problem of multiple speech sources by using two microphones, which has promising applications for auditory scene analysis (ASA) in intelligent service robots. In this paper, we study the DOA estimation based on the inter-channel phase difference (IPD) versus signal frequency framework (IPO-frequency plot). A novel DOA estimation approach has been developed in frequency domain by extracting the high local SNR frequency information aiming to improve the DOA estimation accuracy and the robustness to the noise. The high local SNR frequency information is extracted effectively by exploring the harmonic structure of speech sources in cepstrum domain. The time delay is estimated from the IPDfrequency plot using clustering methods. Experimental results show that the proposed multi-source cepstrumbased DOA estimation algorithm is robust to the additive Gaussian noise, has less complexity and higher DOA estimation accuracy (especially under low SNR condition) compared to those of sinusoidal modeling based DOA estimation method.