Enhanced Double-talk Detection Based on Pseudo-coherence in Stereo

ABSTRACT This paper aims at double-talk detection for acoustic echo cancellers enabling transmission of speech signals in stereo. Double-talk detectors are affected by the same known non-uniqueness problem as echo cancellers. However, echo cancellers are usually adapted using NLMS-like iterative algorithms. Double-talk detectors being based on pseudo-coherence employ traditional spectral estimation techniques, which involve the use of temporal windows. These windows cause a bias of the coherence between the loudspeaker channels: this bias causes the magnitude squared coherence to be smaller than one. Therefore, the non-uniqueness problem known for echo cancellers does not exist for the mentioned type of double-talk detectors. However, the correlation between the loudspeaker channels provoke a strong bias of the inherent estimation of the echo paths. Consequently, we propose a method to decrease the influence of inter-channel correlations on the reliability of the double-talk detector.