If I wanted to label a known speaker by using predefined speaker embeddings (created using pyannote embeddings) could I use these embeddings and create a new centroid at the beginning of the conversation? Or would these embeddings be too specific for the speaker to be identified during the conversation? Especially at the beginning of the conversation as the centroids are few and ambiguous?
Or is the only option to get an accurate identification of the known speaker is at the conclusion of the conversation. By using the average distances between the centroids and the known speaker embeddings to identify the known speaker?
At which point the speaker would be renamed ("speaker 0" becomes "John" for example).

It would be great to init the known speaker at the beginning as it has many advantages including accurate identification and appropriate allocation of speech segments.

What would be the best way to implement this sort of strategy at the beginning of a conversation? Init centroid or something different?