Robust MUSIC Based TDOA Estimation in Competing-Speaker Scenarios

EasyChair Preprint 2317

Abstract

Deep neural network (DNN) based time difference of arrival (TDOA) estimation methods such as Multiple Signal Classification (MUSIC) report superior performance in noisy and reverberation environments but the degradation are observed in the presence of competing for interference. This study investigates its potential for robust MUSIC-based TDOA estimation in competing-Speaker scenarios. First, a time-frequency (TF) mask which is 0 for nonspeech TF bins and 1 for speech TF bins based on the phase and DNN is proposed to accurately estimate the spatial covariance matrix (SCM) that are relatively clean for the MUSIC algorithm in this paper. Second, the proposed approach further reduces the search space to drastically decrease the computation cost by leveraging phase information above. Experimental results on simulated and recorded data confirm the effectiveness and the superiority of the proposed MUSIC-based TDOA estimation method in competing-Speaker scenarios, in comparison with baseline methods.

Keyphrases: DNN, TDOA, music, phase computation cost

@booklet{EasyChair:2317,
  author    = {Md Ahsan Habib and Yi Zhou and Feng Ni},
  title     = {Robust MUSIC Based TDOA Estimation in Competing-Speaker Scenarios},
  howpublished = {EasyChair Preprint 2317},
  year      = {EasyChair, 2020}}