Course detail

Modern Methods of Speech Processing

FIT-MZDAcad. year: 2017/2018

From simple systems to stochastic modelling. Hidden Markov models. Large vocabulary continuous speech recognition. Language models. Speech production, speech perception: time and frequency. Data-driven methods for feature extraction. Speech databases. Excitation in speech coding, CELP. Speaker identification.

Language of instruction

Czech

Number of ECTS credits

0

Mode of study

Not applicable.

Guarantor

prof. Dr. Ing. Jan Černocký

Department

Department of Computer Graphics and Multimedia (UPGM)

Learning outcomes of the course unit

This course allows students to implement simple speech processinga pplications, as for example voice command of a process. However, first of all it enables them to join the development of complex systems for speech recognition and coding systems, using modern methods, in academic and industrial environments.

Prerequisites

basic knowledge of digitial signal processing, having attended a basic course on speech processing is advantageous.

Co-requisites

Not applicable.

Planned learning activities and teaching methods

Not applicable.

Assesment methods and criteria linked to learning outcomes

Study evaluation is based on marks obtained for specified items. Minimimum number of marks to pass is 50.

Course curriculum

Syllabus of lectures:
  1. Review of notions: signal vectors and parameter matrices, basic statistics.
  2. Stochastic modeling of parameters, modeling of time by state sequences.
  3. Hidden Markov models: basic structure, training.
  4. Recognition of speech using HMM: Viterbi search, token passing.
  5. Pronunciation dictionaries and language models.
  6. Speech production and derived parameters: LPC, Log area ratios, line spectral pairs.
  7. Speech perception and derived parameters: Mel-frequency cepstral coefficients, Perceptual linear prediction.
  8. Temporal properties of hearing - RASTA filtering.
  9. Training the feature extractor on the data - linear discriminant analysis.
  10. Speech databases: standards, contents, speakers, annotations.
  11. Vocoders and modeling of the excitation: multi-pulse and stochastic excitations (GSM coding).
  12. CELP coding: long-term predictor, codebooks. Very low bit-rate coders.
  13. Current methods of speaker identification and verification.

Work placements

Not applicable.

Aims

We will mention methods currently implemented in industrial applications (such as mobile phones or commercially available recognizers) but will not promissing methods existing so far only in laboratories. Attention will be paid to techniques derived using data and inspired by human autition and speech production.

Specification of controlled education, way of implementation and compensation for absences

attending the course is not checked, the evaluation of the course is upon the results of exam or final report.

Recommended optional programme components

Not applicable.

Prerequisites and corequisites

Not applicable.

Basic literature

  • Psutka, J.: Komunikace s s počítačem mluvenou řečí. Academia, Praha, 1995
  • Gold, B., Morgan, N.: Speech and audio signal processing, John Wiley & Sons, 2000
  • Texty z http://www.fit.vutbr.cz/~cernocky/speech/

Recommended reading

  • Moore, B.C.J., : An introduction to the psychology of hearing, Academic Press, 1989
  • Jelinek, F.: Statistical Methods for Speech Recognition, MIT Press, 1998
  • Fukunaga, K.: Introduction to Statistical Pattern Recognition, Academic Press, 1990
  • Vapnik, V. N.: Statistical Learning Theory, Wiley-Interscience, 1998
  • Dutoit, T.: An Introduction to Text-To-Speech Synthesis, Kluwer Academic Publishers, 1997

Classification of course in study plans

  • Programme VTI-DR-4 Doctoral

    branch DVI4 , any year of study, winter semester, elective

Type of course unit

 

Lecture

39 hours, optionally

Teacher / Lecturer

prof. Dr. Ing. Jan Černocký

Syllabus

  1. Review of notions: signal vectors and parameter matrices, basic statistics.
  2. Stochastic modeling of parameters, modeling of time by state sequences.
  3. Hidden Markov models: basic structure, training.
  4. Recognition of speech using HMM: Viterbi search, token passing.
  5. Pronunciation dictionaries and language models.
  6. Speech production and derived parameters: LPC, Log area ratios, line spectral pairs.
  7. Speech perception and derived parameters: Mel-frequency cepstral coefficients, Perceptual linear prediction.
  8. Temporal properties of hearing - RASTA filtering.
  9. Training the feature extractor on the data - linear discriminant analysis.
  10. Speech databases: standards, contents, speakers, annotations.
  11. Vocoders and modeling of the excitation: multi-pulse and stochastic excitations (GSM coding).
  12. CELP coding: long-term predictor, codebooks. Very low bit-rate coders.
  13. Current methods of speaker identification and verification.