The proposed knowledge distillation architecture for speaker feature extraction. A speaker representation is extracted by considering acoustic features and AF profiles, the network of shared layers is not only to capture the speaker characteristics from acoustic features, but also to learn the speaker-specific information from AF profiles.