---

This work explores the use of speech knowledge for robust speech recognition by first describing the speech signal through a set of broad speech units, and then conducting a more detailed analysis from these broad units. These units are formed by grouping together parts of the acoustic signal that have similar temporal and spectral characteristics. This work first introduces a novel instantaneous adaptation technique to robustly detect broad classes (BCs) from the input signal using the Extended Baum-Welch (EBW) transformations. Recognition experiments indicate that the EBW method offers a 5% relative improvement compared to typical adaptation approaches. Next, we explore utilizing BC knowledge as a pre-processor for segment-based speech recognition systems. Recognition experiments indicate that utilizing BC knowledge as a pre- processor offers a 14% relative improvement over the baseline recognizer in noisy conditions. Finally, this thesis investigates using BC knowledge for island-driven search. Experiments indicate that the island-driven search strategy results in a 3% improvement in accuracy and also provides faster computation time.

---

Tara Sainath received her PhD from MIT in 2009 and then joined the speech recognition group at IBM. She has organized a special session on sparse representations at Interspeech 2010 and has served as a staff reporter for the IEEE Speech and Language Processing Technical Committee Newsletter. Her main research interests are in acoustic modeling.

---