The EURECOM Submission to the First DIHARD Challenge
The first DIHARD challenge aims to promote speaker diarization research and to foster progress in domain robustness. This paper reports EURECOM's submission to the DIHARD challenge. It is based upon a low-resource, domain-robust binary key approach to speaker modelling. New contributions include the use of an infinite impulse response - constant Q Mel-frequency cepstral coefficient (ICMC) front-end, a clustering selection / stopping criterion algorithm based on spectral clustering and a mechanism to detect single-speaker trials. Experimental results obtained using the standard DIHARD database show that the contributions reported in this paper deliver relative improvements of 39% in terms of the diarization error rate over the baseline algorithm. An absolute DER of 29% on the evaluation set compares favourably with those of competing systems, especially given that the binary key system is highly efficient, running 63 times faster than real-time.
PDF Abstract