Usually cosine similarity of speaker embeddings. Separate from naturalness — a clone can sound good and still be the wrong voice. Connections No outgoing connections.
