•   Closed-loop data-driven modeling    supervised only from a word-level    transcription •   The expectation/maximization (EM)    algorithm is used to improve our    parameter estimates. •   Computationally efficient training    algorithms (Forward-Backward)    have been crucial. •   Batch mode parameter updates are    typically preferred. •   Decision trees are used to optimize    parameter-sharing, system    complexity, and the use of additional    linguistic knowledge.