Return to Main

Introduction:
  Title
  Outline
  Introduction
  Introduction (cont.)
  State of the Art
  State of the Art (cont.)
  Performance Factors
    Noise Environment
    User Population
    Speech Style
    Complexity
    Decade
    Present
    In Five Years

Evaluation Metrics:
  Evolution
  Human Performance
  Machine Performance
    Evolution of Task
  Beyond WER: Named Entity
    Named Entity
    WER
    Beyond WER

Recognition Architectures:
  Why so difficult?
    Overlap
  Theoretic Approach
    Bayesian
    Approach
    Components
  Multiple Knowledge Sources
    Acoustic Front-end
    Acoustic Models
    Language Model
    Search

Acoustic Modeling:
  Feature Extraction
    Measurement
    Spectral Analysis
  Hidden Markov Models
  Parameter Estimation
    Initialization
    Single Gaussian
    Two-Way Split
    Mixture Distribution
    Four-Way Split
    Reestimation
    Optimizing

Language Modeling:
  Wheel of Fortune
  N-Grams
    Bigrams
    Trigrams
  Integration of Natural Language
    Word-level
    Natural Language

Implementation Issues:
  Resource Intensive
    Requirements
  Dynamic Programming-Based Search
    Hypothesis
  Cross-Word Decoding
  Decoding Example
    SENT_START
    WOULD
    GUESS
    GUESS
    EVERY
    REALLY
    SAY
    THING
    SENT_END
  Internet Based Speech Recognition

Technology:
  Conversational Speech
  Indexing of Broadcast News
  Real-Time Translation
    Imagine the Future
    Human Language Engineering
  Future Directions
    Challenges
    Algorithmic Issues

Conclusion and Future Directions:
  References
  Trends
  Limitations on Applications
  Applications on the Horizon
    High Tech heretic
    Beulah Arnott
    BravoBrava

Reading Pal:
  Reading Pal
    Child Reads
    Errors in Red
    Playback
    Word Look-up
    Listen

Summary:
  Goal: Speech Better than Text