Data generation for speaker diarization by speaker transition information

A speaker-turn-aware training data generation method for end-to-end neural diarization (EEND) has been studied. To compensate the lack of real multi-talker training corpus, using artificial multi-speaker conversation data generated from several single speaker’s speech has been a promising way to train EEND models. It is still expected that making the training data close to real multi-talker situation can improve EEND performance. In order to generate a data that contributes to the better performance of trained EEND model, we investigated two speaker alternation handling methods. One is a data-dependent model that introduces a probabiliry of speaker transition extracted from multi-talker data. The other one is a data-independent model that uniformly samples the speaker-alternation probability. Experimental results have showed that the data-dependent method improves the DER as compared with baseline simulated conversations (SC) method on 3-speaker and 4-speaker scenarios. Moreover, it was also confirmed that the data-independent method can achieve comparable performance on the best model on 3 and 4-speaker scenarios.