Detection of overlapping speech in meeting recordings using the modified exponential fitting test

Detection of overlapping speech in meeting recordings is a challenging problem due to both the nature of the conversation itself and the surrounding environment. Accurate identification of these sections of the recording is a crucial first step for speech recognition techniques as their non-detection leads to severe degradation in performance. A possible approach to solving this problem is the use of source number estimation techniques based on the ordered profile of the spatial correlation matrix eigenvalues. In this paper we propose two approaches for detecting overlapping speech based on the Exponential Fitting Test (EFT) a source number estimation technique proposed in [1]. Firstly we propose a frequency domain implementation of the EFT, which is more appropriate when dealing with the broadband speech signals encountered in meetings. We then propose a second approach in which a correction factor is added to allow for the presence of reverberation. The performances of the proposed schemes are evaluated and compared to that of the original EFT using real meeting recordings.