Unsupervised spoken term detection with spoken queries by multi-level acoustic patterns with varying model granularity.