Consideration of The Difference in Accuracy Between Synthetic and Non-Synthetic Speech in Diarization

In recent years, speech segments detection technique called speaker diarization is becoming increasingly important, mainly for meetings, news, telephone speech and so on. However, conventional speaker diarization methods using neural networks require a huge amount of training data. We have shown that speaker diarization is possible with realistic training data by dividing relatively short speech signals recorded individually for each target person and synthesizing them on a computer. In this paper, we investigate the robustness of the method to speech recorded by microphones from multiple sound sources, and also examine the difference in accuracy depending on the position of the sound sources.