Isolated Word ASR

1
Isolated Word ASR

• HMM Training
• Use set of R training observations
• Initialisation
• Viterbi Alignment (Viterbi Training)
• Used by HTKs HInit
• Reestimation
• Baum-Welch (FB) Algorithm
• We will not cover this
• HMM Recognition
• Viterbi

2
Means Variances

• Mean

• Variance

3
HMM Initialisation using Viterbi Training

• Evenly divide observations amongst states
• Calculate means and variances of observations
  associated with each state

4
HMM Initialisation

• Use Viterbi alignment to calculate the most
  likely state sequence
• Using the new observation-state alignment
  recalculate means and variances
• Repeat this process until the estimates do not
  change
• Transition probabilities may, or may not, be
  adjusted

5
Unidimensional Example

• Consider the sequence
• 1, 0, 1, 6, 8, 7, 5, 6, 7, 5, 7, 18, 16, 20, 2,
  0, 1, 0, 2
• The sequence is to be used to Viterbi train a
  4-state left-right model with no skip states.
• Can you do this roughly on paper?
• Could you code the solution?