1. Informatique et Phonétique
Master Plurital
Informatique et
Phonétique

2. Le but de ce cours est de présenter une approche similaire à celle du TAL …. … mais appliquée au signal sonore :
Précédemment ...
1- L'analyse automatique du signal sonore
2- Un langage de requêtes sur le signal sonore
3- Des applications extérieures au signal

3. Cette année ...
TTS
"Text-To-Speech"
J’ai été conçu...
Ma voix...

4. Plan historique de la synthèse Text-To-Speech
synthèse par concaténation de diphones
phonétiseur (graphème  phonème)
Speech Markup languages

5. Forget about the « toothpaste » metaphor! (cf. Allen ’s words)
TTS = NLP + DSP
TEXT
SP EECH
DIGITAL SIGNAL
PROCESSING
Mathematical models
Algorithms
Computations
NATURAL LANGUAGE
Linguistic formalisms
Inference engines
Logical inferences
Phones
Prosody
TEXT-TO-SPEECH SYNTHESIZER
Narrow
Phonetic
Transcription
Forget about the « toothpaste » metaphor! (cf. Allen ’s words)

6. Challenges Intelligibility Naturalness! Engineering :
Coarticulation! Speech synthesis does not reduce to a mere concatenation of recorded phonemes
Segmental quality
Naturalness!
Coarticulation again...
Intonation and phoneme durations must be “coherent”; easy to produce unnatural prosody
Engineering :
low computational and memory cost
Easy adaptation to other languages

7. Omer Dudley’s Voder (Bell Labs, 1936)

8. John Holmes’ formant synthesizer (1964)
Haskins Labs (1968) DecTalk (1983) InfoVox ( )
Rule-based Synthesis

9. Challenges? Intelligibility: more or less... Naturalness : NO
Engineering :
low computational and memory cost : YES
Easy adaptation to other languages : NO

10. Diphone concatenation (1977)al is pS
si~

11. Diphone concatenation (1977)

12. Christian Hamon’s PSOLA (1988)
Cnet (1990) Limsi (Paris , 1992)

13. MBROLA (1993)
Based on the same Poisson’s sum formula as PSOLA, but using edited diphones
Similar overall quality as PSOLA
Same computational load
Completely automatic!  can be used to create lots of compatible synthesizers
J’ai été conçu...
Ma voix...

14. The MBROLA project

15. Challenges? Intelligibility: YES Naturalness : more or less...
Engineering :
low computational and memory cost : more or less...
Easy adaptation to other languages : YES

16. Automatic unit selection
Diphone-based synthesis

17. Automatic unit selection
Unit selection-based synthesis

18. Automatic unit selection
Costs ? Open problem
Concatenation cost ?
Target cost ?
Weights? Trained by resynthesizing the corpus and trying to minimize the difference between original and synthetic
Towards passing the TURING test ?
(ATR, 1996)
(Accuvoice, 1996)
(Univ. Edinburgh, 1997)
(AT&T, 1998)
(L&H, 1999)

19. Towards corpus-based techniques
1995-?: The database years
For automatic phonetization
For automatic generation of intonation and phoneme duration
For automatic selection of units for concatenative synthesis

20. Large text and speech corpora
Tagged text corpora required for training language models for ASR, phonetizers and taggers for TTS
Phonetically labeled speech corpora needed for ASR (multi-speaker, 100s hours) and TTS (single speaker, 1-5 hours)
ELRA (European Language Resource Agency) and LDC (Language Data Consortium) collect and distribute databases
From expert-based systems to corpus-based systems

21. présentation de Mbrola