The speaker embedding is an additional input into the model that captures a particular speaker's voice characteristics.