Audio Segment Retrieval Using a Synthesized HMM