{"id":100885,"date":"2021-08-01T10:23:19","date_gmt":"2021-08-01T08:23:19","guid":{"rendered":"https:\/\/prohoster.info\/blog\/novosti-interneta\/obnovlenie-golosovyh-dannyh-mozilla-common-voice-7-0"},"modified":"2021-08-01T10:23:19","modified_gmt":"2021-08-01T08:23:19","slug":"obnovlenie-golosovyh-dannyh-mozilla-common-voice-7-0","status":"publish","type":"post","link":"https:\/\/prohoster.info\/sq\/blog\/novosti-interneta\/obnovlenie-golosovyh-dannyh-mozilla-common-voice-7-0","title":{"rendered":"P\u00ebrdit\u00ebsimi i t\u00eb dh\u00ebnave t\u00eb z\u00ebrit Mozilla Common Voice 7.0","gt_translate_keys":[{"key":"rendered","format":"text"}]},"content":{"rendered":"
Kompani NVIDIA dhe Mozilla prezantuan p\u00ebrdit\u00ebsimin e grupeve t\u00eb t\u00eb dh\u00ebnave vokale, t\u00eb mbledhura si rezultat i iniciativ\u00ebs Common Voice dhe q\u00eb p\u00ebrfshijn\u00eb shembuj t\u00eb shqiptimit nga 182 mij\u00eb njer\u00ebz, q\u00eb \u00ebsht\u00eb 25% m\u00eb shum\u00eb se 6 muaj m\u00eb par\u00eb. T\u00eb dh\u00ebnat jan\u00eb publikuar si pron\u00eb publike (CC0). Grupet e propozuara mund t\u00eb p\u00ebrdoren n\u00eb sistemet e m\u00ebsimit t\u00eb makinave p\u00ebr nd\u00ebrtimin e modeleve p\u00ebr njohjen dhe sintez\u00ebn e foljes. <\/p>\n
N\u00eb krahasim me p\u00ebrdit\u00ebsimin e kaluar, v\u00ebllimi i materialeve t\u00eb folura n\u00eb koleksion u rrit nga 9 n\u00eb 13.9 mij\u00eb or\u00eb foljeje. Numri i gjuh\u00ebve t\u00eb mb\u00ebshtetura u rrit nga 60 n\u00eb 76, duke p\u00ebrfshir\u00eb p\u00ebr her\u00eb t\u00eb par\u00eb mb\u00ebshtetje p\u00ebr gjuh\u00ebt Belaruse, Kazake, Uzbeke, Bullgare, Armenishte, Azerbajxhanase dhe Bashkire. Grupi p\u00ebr gjuh\u00ebn ruse mbulon 2136 pjes\u00ebmarr\u00ebs dhe 173 or\u00eb materiale foljeje (ishe 1412 pjes\u00ebmarr\u00ebs dhe 111 or\u00eb), nd\u00ebrsa p\u00ebr gjuh\u00ebn ukrainase \u2014 615 pjes\u00ebmarr\u00ebs dhe 66 or\u00eb (ishe 459 pjes\u00ebmarr\u00ebs dhe 30 or\u00eb). <\/p>\n
N\u00eb p\u00ebrgatitjen e materialeve n\u00eb anglisht mor\u00ebn pjes\u00eb m\u00eb shum\u00eb se 75 mij\u00eb njer\u00ebz, duke regjistruar 2637 or\u00eb t\u00eb foljes s\u00eb konfirmuar (ishe 66 mij\u00eb pjes\u00ebmarr\u00ebs dhe 1686 or\u00eb). \u00cbsht\u00eb interesante se gjuha e dyt\u00eb me m\u00eb shum\u00eb t\u00eb dh\u00ebna t\u00eb akumuluara \u00ebsht\u00eb gjuha rwandese, p\u00ebr t\u00eb cil\u00ebn jan\u00eb mbledhur 2260 or\u00eb. Pas saj renditen gjermanishtja (1040), katalanasja (920) dhe esperanto (840). Nga gjuh\u00ebt me rritje m\u00eb dinamike t\u00eb materialeve vokale p\u00ebrmenden gjuha tailandeze (rritje baze 20 her\u00eb, nga 12 n\u00eb 250 or\u00eb), luganda (nga 8 n\u00eb 80 or\u00eb), esperanto (nga 100 n\u00eb 840 or\u00eb) dhe gjuha tamile (nga 24 n\u00eb 220 or\u00eb). <\/p>\n
N\u00eb kuad\u00ebr t\u00eb pjes\u00ebmarrjes s\u00eb saj n\u00eb projektin Common Voice, kompania NVIDIA ka p\u00ebrgatitur modele t\u00eb trajnuara p\u00ebr sistemet e m\u00ebsimit t\u00eb makinave (p\u00ebrkrah\u00ebse PyTorch) bazuar n\u00eb t\u00eb dh\u00ebnat e mbledhura. Modelet shp\u00ebrndahen si pjes\u00eb e mjetit falas dhe t\u00eb hapur NVIDIA NeMo, i cili, p\u00ebr shembull, tashm\u00eb p\u00ebrdoret n\u00eb sh\u00ebrbimet automatizuese t\u00eb z\u00ebrit t\u00eb MTS dhe Sberbank. Modelet jan\u00eb t\u00eb orientuara p\u00ebr p\u00ebrdorimin n\u00eb sistemet e njohjes s\u00eb foljes, sintez\u00ebs s\u00eb foljes dhe p\u00ebrpunimit t\u00eb informacionit n\u00eb gjuh\u00eb natyrore, dhe mund t\u00eb jen\u00eb t\u00eb dobishme p\u00ebr hulumtuesit q\u00eb punojn\u00eb n\u00eb krijimin e sistemeve t\u00eb dialogut me z\u00eb, platformave p\u00ebr transkriptime dhe qendrave t\u00eb thirrjeve automatizore. Ndryshe nga projektet e disponueshme m\u00eb par\u00eb, modelet e publikuara nuk kufizohen n\u00eb njohjen e gjuh\u00ebs angleze dhe mbulojn\u00eb gjuh\u00eb, akcentet dhe forma t\u00eb ndryshme t\u00eb foljes. <\/p>\n
Kujtojm\u00eb se projekti Common Voice ka si q\u00ebllim organizimin e nj\u00eb pune t\u00eb p\u00ebrbashk\u00ebt p\u00ebr mbledhjen e nj\u00eb baze t\u00eb shablloneve t\u00eb z\u00ebrit, e cila merr parasysh gjith\u00eb diversitetin e z\u00ebrave dhe stilit t\u00eb foljes. P\u00ebrdoruesve u ofrohet t\u00eb lexojn\u00eb frazat q\u00eb shfaqen n\u00eb ekran ose t\u00eb vler\u00ebsojn\u00eb cil\u00ebsin\u00eb e t\u00eb dh\u00ebnave t\u00eb shtuar nga p\u00ebrdorues t\u00eb tjer\u00eb. Baza e akumuluar e t\u00eb dh\u00ebnave me regjistrime t\u00eb shqiptimeve t\u00eb frazave t\u00eb zakonshme t\u00eb foljes njer\u00ebzore mund t\u00eb p\u00ebrdoret pa kufizime n\u00eb sistemet e m\u00ebsimit t\u00eb makinave dhe n\u00eb projektet hulumtuese. <\/p>\n
Sip\u00ebrmarr\u00ebsi i bibliotek\u00ebs p\u00ebr njohjen e foljes s\u00eb p\u00ebrzier Vosk beson se disavantazhet e grupit Common Voice jan\u00eb nj\u00ebanshm\u00ebria e materialeve vokale (dominate e njer\u00ebzve t\u00eb seksit meshkuj nga 20-30 vje\u00e7, dhe mungesa e materialeve nga z\u00ebrat e grave, f\u00ebmij\u00ebve dhe t\u00eb moshuarve), mungesa e variacioneve n\u00eb fjalor (rip\u00ebrs\u00ebritja e frazave t\u00eb nj\u00ebjta) dhe shp\u00ebrndarja e regjistrimeve n\u00eb nj\u00eb format q\u00eb shkakton d\u00ebmtime si MP3. \n Burimi: opennet.ru<\/a> <\/p>","protected":false,"gt_translate_keys":[{"key":"rendered","format":"html"}]},"excerpt":{"rendered":"