Speech Synthesis Review

1
Speech Synthesis Review

• Language modelling generates phone sequence and
  prosody of the target utterance.
• Tokenisation
• Parsing helps phrasing
• POS tagging helps disambiguate word senses which
  may have varying phone sequences and prosodies
• Word pronunciation guided by a lexicon (possibly
  subject to morphological preprocessing), with
  non-entries relying on letter to sound rules
• Prosodic modelling is often machine learnt using,
  for example, CART trees

2
Unit Selection Automation in TTS

• Unit selection uses large database of recorded
  units e.g. diphones, phones, syllables, words,
  etc.
• Multiple occurences of the units cover a wide
  space of the spectral and prosodic parameters
• Units nearest in this space to the targets will
  be chosen and will require only minor
  modification
• Database searched for candidates which are ranked
  using a target cost function and concatenative
  cost function

3
Best unit sequence

• Target cost
• Prosodic and spectral closeness to target
• Concatenative cost
• Units occurring beside each other in the recorded
  database being given a zero
• Viterbi algorithm used to find the overall
  minimum cost path.

4
Prosodic Manipulation

• Signal processing is important in the waveform
  generation stage.
• It can help smooth spectral mismatches at unit
  boundaries.
• Also important in duration and F0 manipulation.
• Most common techniques
• Sinusoidal modelling (harmonicsnoise)
• TD-PSOLA
• GELP.