Introduction: LAFF and Vowel Landmarks

Lexical Access From Features (LAFF)

• knowledge based approach to a speech recognition front end
• intended to reflect human perception of speech
• detect landmarks in speech signal
  • alternative to segmental division in time of the speech signal
  • match continuous acoustic signal to discrete lexicon
• attach distinctive features to the landmarks
• Previous work
  • Semivowel analysis [Espy-Wilson 87]
  • Consonant Landmark Detector [Liu 96]
  • Glide Landmark Detector [Sun 96]
  • Lexical matching algorithm [Zhang 98]
  • Voice Feature Detector [Choi 99]

LAFF Landmarks

• Event based
  • Vowel landmarks at the maximum of low frequency energy
  • intervocalic Glide landmarks at the minimum of low frequency energy
  • Consonant landmarks at high frequency discontinuities
• Non-event based "floaters"
  • prevocalic Glide landmarks
  • some Vowels, typically in vowel-vowel sequences

Issues facing robust speech recognition

• Sources of speech variability
• Lexical vs. acoustic representations
• Statistical vs. knowledge based algorithms
• Modular vs. integral architecture

Design goals

• Robustness against speech variability
• Robustness against cascade failure
• Robustness against input conditions - talker, talking rate, level

Optimization criteria

• Receiver Operating Characteristic
• Minimize error rate
• Maximize information output
• Maximize confidence in output