{"id":121971,"date":"2025-02-17T14:22:05","date_gmt":"2025-02-17T12:22:05","guid":{"rendered":"https:\/\/prohoster.info\/blog\/novosti-interneta\/opublikovana-ai-model-sinteza-rechi-zonos-podderzhivayushhaya-klonirovanie-golosa"},"modified":"2025-02-17T14:22:05","modified_gmt":"2025-02-17T12:22:05","slug":"opublikovana-ai-model-sinteza-rechi-zonos-podderzhivayushhaya-klonirovanie-golosa","status":"publish","type":"post","link":"https:\/\/prohoster.info\/sq\/blog\/news\/opublikovana-ai-model-sinteza-rechi-zonos-podderzhivayushhaya-klonirovanie-golosa","title":{"rendered":"Modeli AI p\u00ebr sintez\u00ebn e z\u00ebrit Zonos \u00ebsht\u00eb publikuar, e cila p\u00ebrkrah klonimin e z\u00ebrit","gt_translate_keys":[{"key":"rendered","format":"text"}]},"content":{"rendered":"

Kompania Zyphra publikoi versionin e saj t\u00eb par\u00eb beta t\u00eb modelit AI p\u00ebr sintez\u00ebn e z\u00ebrit Zonos n\u00ebn licenc\u00ebn Apache 2.0. Vegla e ofruar me modelin mb\u00ebshtet funksionin e klonimit t\u00eb z\u00ebrit, duke lejuar q\u00eb z\u00ebri t\u00eb sintetizohet me z\u00ebrin e d\u00ebshiruar, p\u00ebr t\u00eb cilin mjafton t\u00eb ofrohet nj\u00eb regjistrim xhest 30-sekond\u00ebsh t\u00eb z\u00ebrit t\u00eb fol\u00ebsit. Mb\u00ebshtetet sinteza n\u00eb anglisht, japonezisht, kinezisht, francezisht dhe gjermanisht. <\/p>\n

Modeli mbulon 1.6 miliard parametra dhe \u00ebsht\u00eb trajnuar me 200 mij\u00eb or\u00eb regjistrimesh audio. Mb\u00ebshtetet sinteza e z\u00ebrit monotonic (si n\u00eb audiop\u00ebrshtatje) dhe t\u00eb z\u00ebrit emocional (si n\u00eb biseda t\u00eb drejtp\u00ebrdrejta), si dhe sinteza n\u00eb baz\u00eb t\u00eb nj\u00eb prefiksi t\u00eb caktuar (ofrohet nj\u00eb regjistrim audio me fillimin e z\u00ebrit, mbi t\u00eb cilin modeli sintetizon vazhdimin sipas tekstit t\u00eb caktuar, duke ruajtur karakteristikat origjinale t\u00eb z\u00ebrit, p\u00ebr shembull, duke vazhduar t\u00eb flas\u00eb me z\u00eb t\u00eb ul\u00ebt). <\/p>\n

N\u00eb daljen e saj, gjenerohet z\u00eb me frekuenc\u00eb kampionimi 44kHz. Mb\u00ebshtetet z\u00ebvend\u00ebsimi i inserteve t\u00eb sintetizuara p\u00ebr t\u00eb simuluar performanca me shum\u00eb fol\u00ebs ose p\u00ebr ta nd\u00ebrtuar dialogje interaktive, si dhe shtimi i etiketave p\u00ebr t\u00eb kontrolluar shpejt\u00ebsin\u00eb e z\u00ebrit, tonalitetin dhe shprehjen e emocioneve, si g\u00ebzim, frik\u00eb, trishtim dhe zem\u00ebrim. <\/p>\n

Sipas zhvilluesve, n\u00eb cil\u00ebsin\u00eb e z\u00ebrit t\u00eb gjeneruar, modeli nuk \u00ebsht\u00eb m\u00eb i dob\u00ebt ose tejkalon t\u00eb gjitha sistemet publike dhe komerciale t\u00eb sintez\u00ebs q\u00eb jan\u00eb n\u00eb dispozicion (n\u00eb testime ofrohet nj\u00eb krahasim me ElevenLabs, Cartesia dhe FishSpeech). Si disavantazh, sh\u00ebnohet nj\u00eb p\u00ebrqendrim m\u00eb i lart\u00eb i artefakteve audio, si kollitjet, zhurma e frym\u00ebmarrjes ose zhurmat, n\u00eb fillim ose n\u00eb fund t\u00eb materialeve tona audio. <\/p>\n