ASR Design

1
ASR Design
2
ASR Other Issues

• Reading
• HTK Book
• Review Ch.1 Ch.8
• Continuous ASR Token Passing 1.6
• Jurafsky Martin
• Pronunciation Variation 5.7-5.9
• N-grams 6.1-6.6
• ASR Overview 7.1-7.7
• Survey of the State-of-the-Art
• Overview Ch.1

3
Continuous Speech Recognition

• Continuous vs. Connected
• Sub-word models vs. word models
• Join HMM models together using non-emitting
  states
• Need transcription of training data
• Can be signal-aligned or not
• If not alignment achieved automatically using
  Viterbi
• Token Passing
• Like Viterbi keeps track of most likely word
  sequence
• N-best candidate word sequences
• Postprocessed by language model

4
Language Modelling

• Output of recogniser can be constrained by
  language modelling
• Simplest language model No language model
• Next Simplest (?) model P(w)
• Other models
• FSAs
• Weighted automata
• Subword
• Pronununciation variation
• Word sequence probabilities
• N-grams replace P(w)
• (P)CFGs

5
Evaluation

• Word Error Rate

• Spoken corpora
• Uniformity in data for system comparison
• Timit
• Carefully designed collection of sentences
• Switchboard (telephone quality)

6
Future Directions

• Robustness
• Portability
• Adaptation
• Language Modelling
• Out-of-Vocabulary Words
• Spontaneous Speech
• Prosody
• Modelling Dynamics