Automatic Speech Recognition
Lecture notes and assignments on classic speech recognition: acoustics, signal representation, Hidden Markov Models and language modeling.
Building a system that recognizes speech draws on acoustics, statistics and language modeling all at once. This course begins with the acoustic theory of speech production, acoustic-phonetics and signal representation, then turns to the algorithms: pattern classification, search, stochastic modeling and language modeling. Its last part compares approaches and looks at robust recognition, speaker adaptation, paralinguistic information and speech understanding.
What You Will Achieve
- Describe speech sounds using the acoustic theory of speech production
- Build a signal representation and apply vector quantization to it
- Explain how Hidden Markov Models and language models work together in a recognizer
- Compare segment-based and other approaches to speech recognition
Program Overview
Speech and Signals
Background on how speech is produced and represented.
- Acoustic theory of speech production
- Speech sounds
- Signal representation
- Vector quantization
Recognition Algorithms
The core algorithmic methods.
- Pattern classification and search
- Hidden Markov Modeling
- Language modeling
- Graphical models
Approaches and Advanced Techniques
Comparing systems and making them more robust.
- Segment-based ASR and finite-state transducers
- Robust ASR and adaptation
- Speech understanding
- Paralinguistic information
Course Completion
Job Outlook
The acoustic and statistical foundations here remain background for work on speech recognition and language processing systems.
Good for: Students of signal processing or language technology who want the classic foundations of speech recognition, covering the acoustics and the statistical models.
Keep in mind: This is a 2003 course built on Hidden Markov Models and the tools of that time, so it predates current approaches to speech recognition. The labs relied on tool kits and workstations provided in class.
Bottom line: A thorough grounding in the acoustics and classic statistical methods of speech recognition, best read as background to newer work.
Frequently Asked Questions
Does the course include assignments?
Yes. The assignments page lists nine, running through acoustic theory, speech sounds, signal representation, acoustic modeling, language modeling, graphical models, Hidden Markov Models and segment-based ASR.
Which reference books does the course name?
The syllabus lists Spoken Language Processing by Huang, Acero and Hon, books by Jelinek and by Rabiner and Juang on speech recognition, Pattern Classification by Duda, Hart and Stork, and Stevens's Acoustic Phonetics.
What did the term project involve?
Projects typically involved creating and evaluating a speech recognizer along a dimension of interest to the student, using provided tool kits of key recognizer components, so minimal programming skills were needed.



There are no reviews yet.