{"id":103848,"date":"2022-04-22T15:36:52","date_gmt":"2022-04-22T13:36:52","guid":{"rendered":"https:\/\/prohoster.info\/blog\/novosti-interneta\/novyj-vypusk-sistemy-sinteza-rechi-silero"},"modified":"2022-04-22T15:36:52","modified_gmt":"2022-04-22T13:36:52","slug":"novyj-vypusk-sistemy-sinteza-rechi-silero","status":"publish","type":"post","link":"https:\/\/prohoster.info\/sq\/blog\/news\/novyj-vypusk-sistemy-sinteza-rechi-silero","title":{"rendered":"Versioni e reja e sistemit t\u00eb sintez\u00ebs s\u00eb z\u00ebrit Silero","gt_translate_keys":[{"key":"rendered","format":"text"}]},"content":{"rendered":"
Nj\u00eb version i ri publik i sistemit t\u00eb sintez\u00ebs s\u00eb z\u00ebrit Silero Text-to-Speech \u00ebsht\u00eb n\u00eb dispozicion. Projekti \u00ebsht\u00eb fokusuar kryesisht n\u00eb krijimin e nj\u00eb sistemi modern dhe me cil\u00ebsi t\u00eb lart\u00eb t\u00eb sintez\u00ebs s\u00eb z\u00ebrit, q\u00eb nuk \u00ebsht\u00eb m\u00eb i dob\u00ebt se zgjidhjet komerciale nga korporatat dhe e aksesueshme p\u00ebr t\u00eb gjith\u00eb ata q\u00eb duan, pa pasur nevoj\u00eb p\u00ebr pajisje serveri t\u00eb shtrenjta. <\/p>\n
Modelet shp\u00ebrndahen n\u00ebn licenc\u00ebn GNU AGPL, por kompania q\u00eb zhvillon projektin nuk zbulohet mekanizmi i trajnimit t\u00eb modeleve. P\u00ebr t'u ekzekutuar, mund t\u00eb p\u00ebrdoret PyTorch dhe framework-e q\u00eb mb\u00ebshtesin formatin ONNX. Sinteza e z\u00ebrit n\u00eb Silero \u00ebsht\u00eb e bazuar n\u00eb p\u00ebrdorimin e algorithemeve dhe metodave t\u00eb procesimit t\u00eb sinjalit, t\u00eb modifikuara n\u00eb thell\u00ebsi dhe moderne t\u00eb rrjeteve neurale. <\/p>\n
Kjo \u00ebsht\u00eb e v\u00ebrejtur se problemi kryesor i zgjidhjeve moderne t\u00eb rrjeteve neurale p\u00ebr sintez\u00ebn e z\u00ebrit \u00ebsht\u00eb se shpesh ato nuk jan\u00eb t\u00eb aksesueshme p\u00ebrve\u00e7se n\u00eb kuadrin e zgjidhjeve t\u00eb paguara n\u00eb cloud, ndon\u00ebse produktet publike kan\u00eb k\u00ebrkesa t\u00eb larta p\u00ebr pajisje, cil\u00ebsi m\u00eb t\u00eb ul\u00ebt ose nuk jan\u00eb produkte t\u00eb p\u00ebrfunduara dhe gati p\u00ebr p\u00ebrdorim. P\u00ebr shembull, p\u00ebr t\u00eb ekzekutuar pa probleme nj\u00eb nga arkitekturat e reja t\u00eb njohura t\u00eb sintez\u00ebs end-to-end, VITS, n\u00eb modin e sintez\u00ebs (dometh\u00ebn\u00eb, jo p\u00ebr trajnim modelesh), k\u00ebrkohen karta grafike me m\u00eb shum\u00eb se 16 gigabajt VRAM. <\/p>\n
Aktualisht, p\u00ebr versionin m\u00eb t\u00eb fundit t\u00eb sintez\u00ebs jan\u00eb t\u00eb disponueshme publikisht 4 z\u00ebra n\u00eb gjuh\u00ebn ruse, por n\u00eb t\u00eb ardhmen e af\u00ebrt do t\u00eb publikohet versioni tjet\u00ebr me k\u00ebto ndryshime: <\/p>\n
\n
Shpejt\u00ebsia e sintez\u00ebs do t\u00eb rritet edhe me 2-4 her\u00eb;\n
Do t\u00eb p\u00ebrmir\u00ebsohen modelet e sintez\u00ebs p\u00ebr gjuh\u00ebt e vendeve t\u00eb CIS: Kalmyk, Tatar, Uzbek dhe Ukrainase;\n
Do t\u00eb shtohen modelet p\u00ebr gjuh\u00ebt evropiane;\n
Do t\u00eb shtohen modelet p\u00ebr gjuh\u00ebt indiane;\n
Do t\u00eb shtohen modelet p\u00ebr gjuh\u00ebn angleze. <\/ul>\n
Disa nga problemet sistematike, t\u00eb pranishme n\u00eb sintez\u00ebn Silero: <\/p>\n
\n
Ndryshe nga zgjidhjet m\u00eb tradicionale p\u00ebr sintez\u00eb, si RHVoice, sinteza Silero nuk ka integrim me SAPI, klient\u00eb t\u00eb leht\u00eb p\u00ebr t'u instaluar dhe integrime p\u00ebr Windows dhe Android;\n
Shpejt\u00ebsia, megjith\u00ebse \u00ebsht\u00eb jasht\u00ebzakonisht e lart\u00eb p\u00ebr nj\u00eb zgjidhje t\u00eb till\u00eb, mund t\u00eb jet\u00eb e pamjaftueshme p\u00ebr sintez\u00ebn n\u00eb koh\u00eb reale n\u00eb procesor\u00eb t\u00eb dob\u00ebt me cil\u00ebsi t\u00eb lart\u00eb;\n
Zgjidhja p\u00ebr vendosjen automatike t\u00eb akcent\u00ebve nuk trajton omografe (fjal\u00eb si zAmok dhe zamOk) dhe prap\u00eb b\u00ebn gabime, por kjo pap\u00ebrfundim\u00ebsi do t\u00eb korrigjohet n\u00eb rishikimet e ardhshme;\n
Versioni aktual i sintez\u00ebs nuk funksionon n\u00eb procesor\u00ebt pa udh\u00ebzime AVX2 (ose nevojitet t\u00eb ndryshoni ve\u00e7mas cil\u00ebsimet e PyTorch), pasi nj\u00eb nga modulet brenda modelit \u00ebsht\u00eb kquantizuar;\n
Libtorch, i disponuesh\u00ebm p\u00ebr platforma mobile, \u00ebsht\u00eb shum\u00eb m\u00eb i r\u00ebnd\u00eb se ONNX runtime, por versioni ONNX i modelit ende nuk ofrohet. <\/ul>\n