Speech Synthesis: Abstract

1
Speech Synthesis Abstract

• Abstract
• This presentation will touch on the overview of
  Text-to-Speech processing based on rule-based
  approach. Generally, Text-to-Speech system or
  also known as TTS can be divided into three main
  components Text Analysis, Prosody Control and
  Speech Synthesis. Text Analysis will be more on
  normalizing the input text and transcribed the
  text into phonemes. In order the speech to be
  uttered naturally, the prosodic features of the
  synthesized speech need to be controlled. The
  last component which is speech synthesis, is
  responsible for generating the speech signal by
  referring to the rules that map the phonetic
  transcription into sets of acoustical parameters.
  At the end of the presentation, a demo of simple
  malay TTS based on rules will be presented

2
Speech Synthesis

• Speech Synthesis Overview
• Simple Malay TTS
• Conclusion
• Future works

3
Overview History

• - Automatically convert text to speech.
• - History 2
• - since 17th century (mechanical tube)
• - 1960s (speech production theory)
• - 1970s (Application of TTS)
• - 1980s (speech waveform stored speech
• segment) the current trend.

4
Speech production

• Note Resonator 3.

5
TTS Block Diagram
Synthesized speech Source
TTS block diagram 2
Text Dictionary
Text Analysis
Text
Prosody Rules
Prosody Control
Acoustical parameter
Speech Synthesis
6
Text Analysis

• 3 major steps 4
• Structure Analysis Determine the paragraphs,
  sentences, and other structure start and end (
  e.g word, phrases).
• Text normalization Convert written text into
  spoken text.
• Example She arrives at 5.30 pm
• she arrives at five thirty pm

7
Text Analysis (cont.d)

• Text-to-phoneme conversion (phonemisation) Will
  convert each word into phonemes.
• Example
• word daisy
• IPA transcription d e I z i
• ARPAbet transcription d ey z I
• Method of phonemisation 1
• 1. Rules-based
• 2. PbA (Pronunciation by Analog)
  the best for english.
• 3. Neural-Network (NETspeak)
• 4. Machine Learning (Nearest
  neighbour)

8
Prosody Control

• Prosody Refer to the aspects of a sentences
  pronunciation.
• Important Natural sound and convey correct
  meaning of sentence.
• Determine the appropriate prosody sentence by
  considering speech features Pitch (melody),
  timing (rhythm), pausing, speaking rate, focused
  words and many others.
• Example She loves me. ( statement)
• She loves me? ( skeptical )

9
Speech Synthesis

• Finally with the phonemes and prosody information
  are used to produce speech sound.
• Current ways
• Concatenation of recorded human speech
• Formant synthesis use signal processing based
  on knowledge of how phonemes sound and prosody
  affect those phonemes.

10
Malay TTS toy demo

• word
• speech sound
• Note
• 1 Used Klatt rules-based method Top to bottom
  strategy. Java source code can be obtained from
  http//ww.John-Waser.com/TextToSpeech/
• 2 Sample of speech segment concatenation in
  java can be obtained from http//www.javaworld.com
  /javaworld/jw-08-2001/jw-0817-javatalk-p.html

Phonemisation rules1
Phonemisation
Speech database
Speech synthesis 2
11
Example

• Example
• Input word dia (her/him)
• Phonemisation d i y a
• Speech Synthesis d.wav i.wavy.wav
• a.wav
• Output speech sound diya.wav

12
Conclusion

• Observation
• Rules-based need a depth knowledge in phonology
  in that particular language. However, recent
  studies has implement statistics method to
  acquire this knowledge
• Phoneme difficult ( require intensive
  complicated rules and produce unnatural sound)

13
Future works

• Will do investigation on PbA instead of machine
  learning since PbA can give the highest
  percentage of phonemisation 1.
• Avoid rules that needed phonologist expertise
  unless rules can be generated using statistical
  methods.
• Focus on syllables or synthesis unit that retain
  more coarticulation so that speech sound can be
  more natural and rules becomes simpler.

14
Reference

• 1Damper et al. 1997. Evaluating the
  pronunciation component of Text-to-Speech Systems
  for English A Performance Comparison of
  Different Approaches. Computer Speech and
  Language.
• 3Jurafsky, D. Martin, J. H. 2000. Speech and
  Language processing An Introduction to Natural
  Language Processing, Computational Linguistics
  and Speech Recognition.
• 2Shirai K and Abe M (Ed.). 2000. Recent
  Progress in Japanese Speech Synthesis. Gordon and
  Breach Science Publisher.
• 3 http//www.acoustics.hut.fi/slemmett/dippa/ch
  ap2.html
• 4 http//java.sun.com/products/java-media/speech
  /forDevelopers/jsapi-guide/index.html