SPEECH RECOGNITION REQUIRES GOOD
PATTERN RECOGNITION AND SEARCH

• Continuous speech recognition is both a pattern recognition and search problem. Why?
  
  
• The decoding process of a speech recognizer finds the most probable sequence of words given the acoustic and language models. Recall our basic equation for speech recognition:
  
  Search is the process of finding the most probable word sequence:
  
  
  
• The complexity of the search algorithm depends heavily on the nature of the search space, which in turn, depends heavily on the language model constraints (e.g., networks vs. N-grams).
  
  
• Speech recognition typically uses a hierarchical Viterbi beam search for decoding/recognition, and A^* stack decoding for N-best and word graph generation.