Deep Segment Attentive Embedding for Duration Robust Speaker Verification