OUCH: Investigating Limitations of Hidden Markov Models in Speech Recognition

Explore the challenges and limitations of Hidden Markov Models in speech recognition systems through this 45-minute conference talk by Jordan Cohen. Delve into the OUCH (Outing Unfortunate Characteristics of HiddenMarkovModels) project, which investigates the mismatch between model assumptions and actual data in frame-to-frame independence. Learn about the project's approach to creating speech data that satisfies model assumptions, and gain insights into the implications of these assumptions on recognition performance. Discover the preliminary findings of a survey conducted among researchers and engineers in the field of speech and language technology. Examine topics such as acoustic and language models, modeling mismatches, language model weight, and active performance. Gain valuable perspectives on the current state of speech recognition technology and potential areas for improvement.

Introduction
OUCH
Acoustic Model
Language Model
Results
Modeling Mismatch
Language Model Weight
Active Performance
Literature Search
Snowball Sampling
Questionnaire
Demographics
Improvements
Snowball
Youngsters
Selfclassifications
brittleness
pronunciation
models dont work
categorization
the hidden agenda