11-756 THEORY AND PRACTICE OF SPEECH RECOGNITION SYSTEMS

THEORY AND PRACTICE OF SPEECH RECOGNITION SYSTEMS

Instructor: Bhiksha Raj

COURSE NUMBER--ECE: 18799D LTI: 11756
Credits:12
Timings:4:30 p.m. -- 5:50 p.m.
Days:Mondays and Wednesdays
Location: GHC 4102

Prerequisites:
Mandatory:  Linear Algebra. Basic Probability Theory.
Recommended:  Signal Processing.
Coding Skills:  This course will require significant programming form the students. Students must be able to program fluently in at least one language (C, C++, Java, Python, LISP, Matlab are all acceptable).


PROJECTS PAGE

Voice recognition systems invoke concepts from a variety of fields including speech production, algebra, probability and statistics, information theory, linguistics, and various aspects of computer science. Voice recognition has therefore largely been viewed as an advanced science, typically meant for students and researchers who possess the requisite background and motivation.

In this course we take an alternative approach. We present voice recognition systems through the perspective of a novice. Beginning from the very simple problem of matching two strings, we present the algorithms and techniques as a series of intuitive and logical increments, until we arrive at a fully functional continuous speech recognition system.

Following the philosophy that the best way to understand a topic is to work on it, the course will be project oriented, combining formal lectures with required hands-on work. Students will be required to work on a series of projects of increasing complexity. Each project will build on the previous project, such that the incremental complexity of projects will be minimal and eminently doable. At the end of the course, merely by completing the series of projects students would have built their own fully-functional speech recognition systems.

In this edition of the course we will also introduce the theory of Weighted Finite State transducers. In the latter half of the course students will learn to build their own WFST systems, and use open-source tools to compose their own WFST recoginzers.

Grading will be based on project completion and presentation.


                                                                                                                        
Class 120 Jan 2014 Introduction Slides
Class 222 Jan 2014 Data capture Slides assignment 1
Class 327 Jan 2014 No class
Class 429 Jan 2014 Feature computation Slides
Class 53 Feb 2014 Assignment 1 presentations
Class 65 Feb 2014 String Matching Slides assignment 2
Class 710 Feb 2014 DTW Slides
Class 812 Feb 2014 assignment 3
Class 917 Feb 2014 Assignment 2 presentations  
Class 1019 Feb 2014 DTW to HMMs Slides
Class 1124 Feb 2014 HMMs, part 1 Slides
Class 1226 Feb 2014 Assignment 3 presentations assignment 4
Class 133 Mar 2014 HMMs, part 2 Slides
Class 145 Mar 2014 Recognizing continuous speech Slides
Class 1517 Mar 2014 Grammars Slides
Class 1619 Mar 2014 Assignment 4 presentations assignment 5
Class 1724 Mar 2014 Backpointer Tables Slides
Class 1826 Mar 2014 Training from continous speech Slides
Class 192 Apr 2014 Ngram Models Slides assignment 6