LING124 Statistical approach to ASR: an overview

1
LING124 Statistical approach to ASR an overview

• October 21, 2008

2
Class outline

• Noisy channel model
• Bayesian inference
• Likelihood acoustic model
• Prior language model
• Schematic architecture

3
Noisy channel model
4
Bayesian inference (1)

• Take every possible sentence in the language
• Run each sentence through the channel and see if
  it matches the original noisy output
• Pick the sentence whose channel output best
  matches the original output

5
Bayesian inference (2)

• Oo1, o2, o3, ... , oT
• Ww1, w2, w3, ... , wn
• L set of all possible word sequences

6
Bayes rule
7
Bayesian inference (3)
8
Likelihood acoustic model (1)

• P(OW)
• Probability of the observation given word
• Observation is typically continuous, rather than
  discrete, in statistical models of ASR
• e.g. MFCC, delta, double-delta
• Probability of a continuous random variable can
  be calculated if we have a probability
  distribution function (pdf)

9
Likelihood acoustic model (2)

• Gaussian distribution is the most often used pdf
  (p307)
• Most simply, one Gaussian distribution for each
  word (or some other linguistic unit)
• Recall that a typical feature vector has 39
  components (MFCC, delta, double-delta)
• We use a multivariate Gaussian distribution for
  each word (or some other linguistic unit)

10
Prior language model

• P(W)
• Probability of observing a word sequence W in the
  given language
• P(w1, w2, ..., wn)
• By chain rule,
• Most often calculated using N-grams

11
Schematic architecture