WORD GRAPH GENERATION AND RESCORING

• An N-best list can be viewed as a graph:
  
  
  This representation is often referred to as a lattice.
  
  
• How might we generate such a graph using time synchronous Viterbi decoding?
  
  

Solution: keep multiple choices at each node in the graph during the dynamic programming step: It is hard to underestimate the impact word graph rescoring has had on speech recognition research. However. these graphs are very large and take at least an order of magnitude more time to generate (than the one-best choice). Why are these worth the trouble? Word graphs can be very large: 10 to 50 MBytes per file; 1 Gbyte or more per corpus.

• What figure of merit can we use to describe such graphs? (Hint: lattice word error rate) Explain the significance of this measure.