Hidden Markov Models for Speech Recognition

CSE552/652
Hidden Markov Models for Speech Recognition

Spring 2006
Mondays/Wednesdays, 12:00 p.m.- 1:30 p.m., Room WCC 403, Wilson Clark Center

John-Paul Hosom
'hosom' at cslu . ogi .edu
503.748.1456

Course Description
Hidden Markov Model-based technology is widely used in today's speech recognition systems. This course is an introduction to the theory and practice of speech recognition using HMM technology. Topics include dynamic time warping, Markov Models and Hidden Markov Models (discrete, semi-continuous, and continuous), vector quantization, Gaussian Mixture Models, the Viterbi search algorithm, the forward-backward training algorithm, language modeling, and speech-specific adaptations of HMMs. The course is focused on understanding these fundamental technologies and developing the main components of speech recognition systems. Students can expect to come away from the course with an ability to write programs for the training and execution of simple HMM systems, and to know how to extend these systems to more complex cases. Prerequisite: C programming experience.

Textbooks
There are two recommended textbooks:

Fundamentals of Speech Recognition
Lawrence Rabiner and Biing-Hwang Juang
Prentice Hall, New Jersey, 1993

Statistical Methods for Speech Recognition
Frederick Jelinek
The MIT Press, Cambridge, MA, 1999

Grading Policy
Grading is based on three programming assignments, a midterm, and a final. The programming projects provide a template for basic functions such as file I/O and the basic program structure; the student must write the relevant functions. The three projects are worth 15%, 20%, and 25% of the total grade, respectively. The midterm is worth 20%, and the final is worth 20%.

Weekly Schedule
Lecture notes are added as links to power-point files. Files related to programming assignments will also be posted here as ZIP files.

Week Number	Links	Lecture Topics
Week 1: April 3, April 5	lecture1 lecture2 project1	Course Overview Why Is Automatic Speech Recognition Difficult? Background: Speech Production, Representations of Speech, Models of Human Speech Recognition General Issues in Developing ASR Systems Induction DTW Motivation / Algorithm / Implementation DTW Examples Assign Project 1 on April 5th.
Week 2: April 10, April 12	lecture3 lecture4	(Review) Relevant Probability / Statistics Background Markov Models Log-Domain Mathematics Hidden Markov Models HMM Topologies Vector Quantization
Week 3: April 17, April 19	lecture5 lecture6	Gaussian Mixture Models Speech Features (LPC, PLP, MFCC) HMMs for Speech
Week 4: April 24, April 26	lecture7 lecture8 project2	Project 1 Due April 24th at midnight Viterbi Search Lots of Viterbi Search Examples Assign Project 2
Week 5: May 1, May 3	lecture9 lecture10	Review Project 1 on May 1 Semi-Markov Models Initializing an HMM Forward Procedure Backward Procedure
Week 6: May 8, May 10	midterm sample lecture11	In-Class Midterm on May 8 Go over Midterm on May 10 Gamma Xi Baum-Welch or Forward-Backward or EM Algorithm for training HMMs Training on Multiple Files
Week 7: May 15, May 17	lecture12 lecture13 project3	Project 2 Due May 15 at midnight Expectation-Maximization in General Embedded Training Search Algorithms: Two-Level Search Algorithms: Level-Building Assign Project 3 on May 17
Week 8: May 22a May22b, May 24	lecture14 lecture15 bonus project	May 22 is a Double Class! Starts at 10:30, goes to 1:30. Review Project 2 on May 22 Search Algorithms: One-Pass Search Strategies: Beam Search, Grammar/Tree Search, On-Line Processing, Balancing Insertions and Deletions, N-Best Output Acoustic-Model Strategies: Semi-Continuous HMMs, State Tying, State Clustering, Cloning, Pause Models
Week 9: May 31a, May 31b	lecture16 lecture17	May 29 is Memorial Day; no class May 31 is a Double Class! Starts at 10:30, goes to 1:30 Language Models: Incorporating N-Gram LM, Linear Smoothing, Good-Turing Smoothing, Discounting and Back-Off, Cache LM, Class-Based LM, Perplexity Other Approaches to ASR: Segment-Based Systems, HMM/ANN Hybrids, others Tree-Based Search with Language Models Evaluation of System Performance and State of the Art
Week 10: June 5, June 7		NO CLASS JUNE 5 OR JUNE 7 (Paul leaves June 5th and returns June 13th)
Finals Week		Project 3 Due June 16 at midnight Optional Project, if done, is due June 16 Take-Home Final Exam Due June 16

Previous Year's Lecture Notes and Project Files

lecture1	lecture2	lecture3	lecture4	lecture5	lecture6
lecture7	lecture8	lecture9	lecture10	lecture11	lecture12
lecture13	lecture14	lecture15	lecture16
project1	project2	project3

CSE552/652 Hidden Markov Models for Speech Recognition

CSE552/652
Hidden Markov Models for Speech Recognition