ENGINEERING TRIPOS PART IIB – 2012/2013

Module 4F11 - Speech and Language Processing

AIMS

The module aims to introduce major techniques for recognising and synthesising speech signals, and for statistical machine translation.

LECTURE SYLLABUS (Prof. P C Woodland and Dr. W J Byrne)

Lecture 1: Overview/Introduction

• Speech production mechanisms, types of speech sound, source-filter model, applications of speech and text processing

Lectures 2-3: Acoustic Analysis

• FFT based methods, Mel scale, cepstral analysis, all-pole filter models, calculation of LP coefficients. formant and voicing analysis. Front-end analysis for speech recognition.

Lectures 4-5: ASR Introduction and Isolated Word Recognition

• Statistical speech recognition, task complexity. Hidden Markov models. Continuous density HMM parameter estimation, Baum-Welch algorithm, Viterbi algorithm, Gaussian mixture models for HMMs.

Lecture 6: Sub-word Acoustic Models

• Large vocabulary speech recogntion, limitations of word models, context dependent phones, parameter tying.

Lecture 7: Language Models

• Perplexity, N-gram language models, discounting, interpolation.

Lecture 8: ASR Search Issues

• Continuous speech recognition. Pruning. Integrating context dependent HMMs and N-gram language models.

Lectures 9-10: Weighted Finite State Transducers for Speech and Language Processing

• Efficient realization of probabilistic models for sequence processing. Transduction, composition, determinization, minimum-cost search. WFSTs in ASR search and other language processing applications.

Lecture 11: Introduction to Statistical Machine Translation

• Statistical pattern processing approaches to translation. Automatic evaluation of translation quality.

Lecture 12: SMT - Alignment

• Parallel text as training data. Models of word and phrase alignment in translation. Model estimation procedures.

Lecture 13: SMT - Translation

• Phrase-based translation systems. Implementation via WFSTs.

Lecture 14: Text-to-Speech Synthesis

• Introduction to TTS. Data-driven synthesis and Hidden Markov Model approaches to TTS.

OBJECTIVES

On completion of the module students should:

• Understand techniques for speech analysis, recognition and synthesis;
• Understand techniques for statistical machine translation;
• Be able to apply these techniques in order to build simple speech recognition, synthesis and machine translation systems;
• Be aware of the current state-of-the-art in speech recognition and synthesis and in machine translation technology.

REFERENCES

Please see the Booklist for Group F Courses for references for this module.

Last updated: June 2012