Unsupervised discovery of linguistic structure including two-level acoustic patterns using three cascaded stages of iterative optimization.