On the Importance of Efficient Transition Modeling for Speaker Diarization

In recent years speaker diarization becomes an important issue. In previous works, we presented the Hidden Distortion Model (HDM) approach, in order to overcome the limitations of traditional HMMs in terms of emission and transition modeling. In this work, we show that HDM allows to build more efficient speaker diarization systems both in terms of diarization error rated and in terms of memory footprint. The best diarization performance is obtained using smaller than usual emission models which constitutes potentially a key advantage for embedded applications with limited memory resources and computational power. A significant memory size reduction was observed using LDC CALLHOME (American) for both SOM- and GMM-based emission probability models.