Speaker Localization in Smartphones using Adaptive Eigenvalue Decomposition with Noise Reduction

Abstract

Most smartphones are dual microphone devices capable of determining the direction of arrival of an utterance from a speaker source. The widespread use of such devices helps in improving hearing aid systems without increased expenses. These types of sound source localization (SSL) systems with two sensors take advantage of time delay estimation (TDE) techniques such as cross-correlation and adaptive eigenvalue decomposition (AED). The former lacks reliability in situations with reverb, while the latter suffers from background noise. In this paper, we observed the effect of integrating a noise reduction algorithm to AED for SSL applications. Given the robustness of AED with room reverb, we expect performance improvement of TDE from noise-reduced outputs. The minimum mean-square error with decision-directed (MMSE-DD) noise estimation algorithm acts as a filter for the received signals. We proposed $\text{MMSE}-\text{DD}+\text{AED}$ to obtain an SSL algorithm in poor environment conditions. The empirical results of the system yielded 69.87%, which is a significant improvement from previous SSL algorithms in smartphones. Furthermore, a tilt compensation solution boosted the accuracy to 79.28%, addressing the dynamic behavior of the built-in microphones of the device.