6.345/HST.728  Automatic Speech Recognition

This course introduces students to the rapidly developing field of automatic speech recognition and spoken language processing. Its content is divided into four parts. Part I deals with the fundamentals of speech including signal representation, acoustic theory of speech production, and acoustic-phonetics. Part II deals with the basics of acoustic modeling and search including vector quantization, Gaussian mixtures, dynamic time-warping, and search. Part III deals with fundamentals of lexical access including hidden Markov models, language modeling, and finite state transducers. Part IV deals with more advanced methods including adaptation, discriminative techniques, and neural-network based modeling. Lectures will be interspersed with theory and applications.

There will be two 90 minute lectures per week, along with office hours. Lecture material will be front-loaded into the first half of the class, so that students are prepared for the term project which will start after spring break. There will be invited guest lectures on related speech and language topics in the latter half of the class.

There will be five assignments interspersed over the first half of the class as well, which include problems, a laboratory, and a brief in-class test. The assignments will be closely linked with lecture material. The laboratories will involve speech recognition experiments that can be performed on student Athena accounts. All material will be made available on the MIT Stellar course website.

Each of the five assignments will count 10% towards the final grade. The final term project will count towards 50% of the final grade.

The particular topics covered in the first half of the class include: digital signal processing, acoustic theory of speech production, acoustic phonetic properties of speech sounds, vector quantization, Gaussian mixture models, dynamic time warping, Viterbi and A* search, hidden Markov models, language modeling, finite state transducers, speaker adaptation, artificial neural networks, sequential training, and Neural End-to-End ASR