{"id":105165,"date":"2022-09-25T21:36:38","date_gmt":"2022-09-25T19:36:39","guid":{"rendered":"https:\/\/prohoster.info\/blog\/novosti-interneta\/otkryt-kod-sistemy-raspoznavaniya-i-perevoda-rechi-whisper"},"modified":"2022-09-25T21:36:38","modified_gmt":"2022-09-25T19:36:39","slug":"otkryt-kod-sistemy-raspoznavaniya-i-perevoda-rechi-whisper","status":"publish","type":"post","link":"https:\/\/prohoster.info\/sq\/blog\/novosti-interneta\/otkryt-kod-sistemy-raspoznavaniya-i-perevoda-rechi-whisper","title":{"rendered":"U hap kodi i sistemit t\u00eb njohjes dhe p\u00ebrkthimit t\u00eb z\u00ebrit Whisper","gt_translate_keys":[{"key":"rendered","format":"text"}]},"content":{"rendered":"
Projekti OpenAI, i angazhuar n\u00eb zhvillimin e projekteve t\u00eb hapura n\u00eb fush\u00ebn e inteligjenc\u00ebs artificiale, publikoi rezultatet e lidhura me sistemin e njohjes s\u00eb z\u00ebrit Whisper. Thuhet se p\u00ebr t\u00eb folur n\u00eb gjuh\u00ebn angleze, sistemi ofron nivele besueshm\u00ebrie dhe sakt\u00ebsie n\u00eb njohjen automatike q\u00eb jan\u00eb t\u00eb ngjashme me njohjen njer\u00ebzore. Kodi i zbatimit t\u00eb referenc\u00ebs \u00ebsht\u00eb i hapur, i nd\u00ebrtuar mbi korniz\u00ebn PyTorch dhe nj\u00eb grup modelet t\u00eb trajnuara tashm\u00eb, t\u00eb gatshme p\u00ebr p\u00ebrdorim. Kodi \u00ebsht\u00eb i hapur n\u00ebn licenc\u00ebn MIT. <\/p>\n
P\u00ebr t\u00eb trajnuar modelin, jan\u00eb p\u00ebrdorur 680 mij\u00eb or\u00eb t\u00eb dh\u00ebnash t\u00eb z\u00ebrit, t\u00eb mbledhura nga disa koleksione q\u00eb mbulojn\u00eb gjuh\u00eb dhe fusha tematike t\u00eb ndryshme. Af\u00ebrsisht 1\/3 e t\u00eb dh\u00ebnave t\u00eb z\u00ebrit t\u00eb p\u00ebrdorura n\u00eb trajnim i p\u00ebrkasin gjuh\u00ebve q\u00eb nuk jan\u00eb angleze. Sistemi i propozuar p\u00ebrpunon sakt\u00ebsisht situata t\u00eb tilla si shqiptimi me theks, prania e zhurmave n\u00eb sfond dhe p\u00ebrdorimi i zhargoneve teknike. P\u00ebrve\u00e7 transkriptimit t\u00eb z\u00ebrit n\u00eb tekst, sistemi gjithashtu mund t\u00eb p\u00ebrkthej\u00eb z\u00ebrin nga \u00e7do gjuh\u00eb n\u00eb anglisht dhe t\u00eb identifikoj\u00eb shfaqjen e z\u00ebrit n\u00eb rrjedh\u00ebn e z\u00ebrit. <\/p>\n
Modelet jan\u00eb formuluar n\u00eb dy p\u00ebrfaq\u00ebsime: nj\u00eb model p\u00ebr gjuh\u00ebn angleze dhe nj\u00eb model shum\u00ebgjuh\u00ebsor, q\u00eb mb\u00ebshtet gjithashtu gjuh\u00ebt ruse, ukrainase dhe belorusisht. Nga ana tjet\u00ebr, \u00e7do p\u00ebrfaq\u00ebsim ndahet n\u00eb 5 variante, q\u00eb dallojn\u00eb n\u00eb madh\u00ebsi dhe numrin e parametrave t\u00eb mbuluar n\u00eb model. Sa m\u00eb e madhe t\u00eb jet\u00eb madh\u00ebsia, aq m\u00eb e lart\u00eb \u00ebsht\u00eb sakt\u00ebsia dhe cil\u00ebsia e njohjes, por edhe k\u00ebrkesat p\u00ebr dimensionin e memories video t\u00eb GPU jan\u00eb m\u00eb t\u00eb larta dhe performanca m\u00eb e ul\u00ebt. P\u00ebr shembull, varianti minimal p\u00ebrfshin 39 milion parametra dhe k\u00ebrkon 1 GB memories video, nd\u00ebrsa varianti maksimal p\u00ebrfshin 1550 milion parametra dhe k\u00ebrkon 10 GB memories video. Varianti minimal \u00ebsht\u00eb 32 her\u00eb m\u00eb i shpejt\u00eb se maksimal.
<\/center> <\/p>\n
Sistemi p\u00ebrdor arkitektur\u00ebn e rrjetit nervor \"Transformer\", e cila p\u00ebrfshin nj\u00eb kodues dhe nj\u00eb dekodues q\u00eb lidhen me nj\u00ebri-tjetrin. Tingulli ndahen n\u00eb fragmente 30 sekond\u00ebshe, t\u00eb cilat konvertohen n\u00eb log-Mel spektrogram dhe i kalohen koduesit. Rezultati i pun\u00ebs s\u00eb koduesit d\u00ebrgohet n\u00eb dekodues, i cili parashikon nj\u00eb p\u00ebrfaq\u00ebsim tekstual, t\u00eb p\u00ebrzier me tokene speciale q\u00eb lejojn\u00eb zgjidhjen e detyrave t\u00eb tilla si p\u00ebrcaktimi i gjuh\u00ebs, mbajtja mend e radh\u00ebs s\u00eb shqiptimit t\u00eb frazave, transkriptimi i fjal\u00ebs n\u00eb gjuh\u00eb t\u00eb ndryshme dhe p\u00ebrkthimi n\u00eb anglisht. \n Burimi: opennet.ru<\/a> <\/p>","protected":false,"gt_translate_keys":[{"key":"rendered","format":"html"}]},"excerpt":{"rendered":"
![\"U](\"\/wp-content\/uploads\/2022\/09\/d6aa63e77cddead1a7abfa9083ff37cb.jpg\")
<\/center> <\/p>\n