Unimodal Aggregation for CTC-based Speech Recognition