Speaker: Juan Ignacio Álvarez Trejos. Abstract: This presentation covers the work presented at Odyssey 2024, focusing on speaker diarization in two-speaker scenarios. End-to-end neural speaker diarization systems are designed to handle overlapping speech while accurately distinguishing between speakers. In this work, we explore integrating speaker embeddings into these systems to improve their speaker discriminative capabilities. We propose several methods to combine speaker embeddings with acoustic features and analyze key aspects such as handling silence frames, selecting the window length for extracting speaker embeddings, and optimizing the transformer encoder size. This approach is evaluated in the context of two-speaker diarization tasks.
