Główne pojęcia
The author utilizes the TitaNet model trained with additive angular margin (AAM) loss to optimize cosine distance between speaker embeddings, using cosine similarity as the back-end metric.
Streszczenie
The TitaNet model is trained end-to-end with AAM loss to enhance speaker embeddings' cosine distance. The paper focuses on verification and diarization experiments utilizing cosine similarity. The formula for optimization is detailed in the content.
Statystyki
The TitaNet model was trained end-to-end with additive angular margin (AAM) loss [19].
For all experiments, cosine similarity is used as the back-end metric.