Statistics for Speaker Diarization Using Improved SincNet Models to Extract Speaker Embeddings