MAGE Schemaclick to enlarge

MAGE logoSta­tis­ti­cal para­met­ric speech syn­the­sis, based on Hid­den Markov Mod­els has been demon­strated to be very effec­tive in syn­the­siz­ing high qual­ity, nat­ural and expres­sive speech. This tech­nique is also able to pro­vide high flex­i­bil­ity as a speech pro­duc­tion model and a small data­base foot­print.
Based on the exist­ing HTS engine, we devel­oped a stream­ing archi­tec­ture of the sys­tem, called performative-HTS or pHTS. On top of pHTS we devel­oped MAGE, a thread safe and engine inde­pen­dent layer of pHTS, that can be used in reac­tive speech syn­the­sis designs, (i.e. a design that can be often inter­rupted and can respond in real-time to requests).
Quan­ti­ta­tive eval­u­a­tions of the sys­tem show that the degra­da­tion of speech qual­ity in pHTS is small with ref­er­ence to HTS, even though pHTS has a delay of one pho­netic label only . These results are sup­ported by a sub­jec­tive and an objec­tive eval­u­a­tion, which con­firms that HTS and pHTS result­ing speech wave­forms can hardly be distinguished.

For more details on the archi­tec­ture of the systems :

RT-HTS nume­di­art report
sHTS p3s paper :: “sHTS: A Stream­ing Archi­tec­ture for Sta­tis­ti­cal Para­met­ric Speech Synthesis”




MAGE pHTS is released under GPL license.
Feel free to down­load it!
MAGE web­site
github iconProject on GitHub


If you have ques­tions or remarks, please con­tact Maria Astri­naki.


MAGE and pHTS has been devel­oped by sev­eral mem­bers of  Uni­ver­sity of Mons – NUMEDIART Insti­tute and Acapela Group :

Maria Astri­naki (Main Maintainer/Developer)
Onur Baba­can
Geof­frey Wil­fart
Alexis Moinet
Nico­las d’Alessandro
Thierry Dutoit