Isolated Word ASR

1
Isolated Word ASR

• HMM Training
• Use set of R training observations
• Initialisation
• Viterbi Alignment
• Reestimation
• Baum-Welch (FB) Algorithm
• HMM Recognition
• Viterbi

2
HMM Initialisation

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

3
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

4
HMM Reestimation

• We wish to estimate transition probabilities A
  and emission probabilities B.
• The initialisation gives a hard assignment of
  training observations to states
• Instead, we should assign each observation to
  each state in proportion to the probability of
  being in state sj at time t when the feature
  vector is observed
• So use rather than

5
HMM Reestimation

• Remember

6
HMM Reestimation

• Means and variances of observations associated
  with each state reestimated using

7
(No Transcript)
8
(No Transcript)
9
Reestimation of A
10
Reestimation of A
If we sum over time t, we get the
expected number of times that state si is
visited. If we sum over time t, we get the
expected number of transitions from state si to
state sj, (when we exclude tT)
11
(No Transcript)
12
(No Transcript)
13
(No Transcript)
14
(No Transcript)
15
(No Transcript)