Special Topics: Large Vocabulary Recognition Course No: 506/606, Winter 2012 Izhak Shafran The Center for Spoken Language Understanding The Oregon Health & Science University Course Information Lectures References Schedule Links

Intimity by Jan Koblasa, image from Statistical Methods for Speech Recognition, with permission from Frederick Jelinek, MIT Press

Course Information

• Introductory lectures
• Reading discussions: Students will take turns presenting papers and will be responsible for up to 2 papers.
• Course project: There will be no final exam. Instead, the course requires a final project of interest to student, chosen in consultation with the instructor. The project requires a written report and a final presentation. In many cases, the data, software toolkit, and key components for the project will be made available.

Lectures and Readings

• Session 1 (1/12)
  
  
  • Basic overview of speech and speech recognition
  • Chapter 2 from MGSY and Lecture 2 & 3 from MITOCW
  
  
• Session 2 (1/19)
  
  
  • The Search Problem: Resorting to Finite State Transducers
  • Search graph using finite state transducers Mohri et al
  • Exact language model Roark et al
  • Context dependency transducer Schuster et al
  • On the fly composition Mc Donough et al
  • Exact lattice using WFST Povey et al
  
  
• Session 3 (2/2)
  
  
  • The Model Estimation Problem: Maximizing Likelihood
    • Chapter 3 from MGSY
    • Heteroscedastic discriminant analysis and reduced rank HMMs for improved speech recognition
  
  
• Session 4 (2/9)
  
  
  • The Model Estimation Problem: Maximizing Discrimination
    • Chapter 4 from MGSY
    • Baum-Eagen on Rational Functions
    • Boosted MMI
    • Feature-space MPE
  
  
• Session 5 (2/16)
  
  
  • The Model Estimation Problem: Other Discriminative Estimations
    • Large-Margin HMM Estimation
    • Large-Margin Online Learning in HMMs
    • Direct Error Rate Minimization of HMMs
  
  
• Session 6 (2/23)
  
  
  • Domain Adaptation: Adapting Acoustic and Language Models
    • Language model adaptation with MAP estimation and the perceptron algorithm
    • Unsupervised language model adaptation
    • Adaptation of maximum entropy capitalizer: Little data can help a lot
    • A Literature Review of Domain Adaptation with Unlabeled Data
  
  
• Session 7 (3/1)
  
  
  • Speaker Adaptation: Adapting Acoustic Models
    • Chapter 5 from MGSY
    • MAP adaptation, Jean Luc-Gauvain et al
    • MLLR adaptation, Leggetter et al
    • Eigenvoices, Kuhn et al
    • Rapid adaptation, CLSP workshop report, 1998
  
  
• Session 8 (3/8)
  
  
  • Who Said When? The Diarization Problem
    • The ICSI RT07s Speaker Diarization System
    • An overview of automatic speaker diarization systems
  
  
• Session 9 (3/15)
  
  
  • Problems in Health and Medical Domain
    • Speech in Parkinson's Disease
    • Neurolinguistic features of spontaneous language production dissociate three forms of neurodegenerative disease: Alzheimer's, Huntington's, and Parkinson's
    • Automatically derived spoken language markers for detecting MCI
    • Gapped BLAST and PSI-BLAST: a new generation of protein database search programs, Altschul SF, Madden TL, SchafferAAet al., Nucleic Acids Research 25:3389-3402, 1997.
    • Protein homology detection by HMM-HMM comparison, Soding J., Bioinformatics 21 (7): 951-960, 2005.
  
  
• Session 10 (3/22)
  
  
  • Course project related topics
  
  

References

• MGSY The Application of Hidden Markov Models in Speech Recognition, Mark Gales and Steve Young
• MITOCW Lecture Notes from Jim Glass' Automatic Speech Recognition Course, MIT OpenCourseWare

Schedule

Links