{"id":53722,"date":"2019-12-08T00:00:00","date_gmt":"2019-12-07T21:00:00","guid":{"rendered":"https:\/\/prohoster.info\/blog\/blog_prohoster\/kompaniya-mozilla-predstavila-dvizhok-raspoznavaniya-rechi-deepspeech-0-6"},"modified":"2020-02-18T14:01:39","modified_gmt":"2020-02-18T11:01:39","slug":"kompaniya-mozilla-predstavila-dvizhok-raspoznavaniya-rechi-deepspeech-0-6","status":"publish","type":"post","link":"https:\/\/prohoster.info\/sq\/blog\/novosti-interneta\/kompaniya-mozilla-predstavila-dvizhok-raspoznavaniya-rechi-deepspeech-0-6","title":{"rendered":"Kompania Mozilla lan\u00e7oi motorin e njohjes s\u00eb gjuh\u00ebs DeepSpeech 0.6","gt_translate_keys":[{"key":"rendered","format":"text"}]},"content":{"rendered":"

T\u00eb paraqitura<\/a><\/noindex> publikimi i motorit t\u00eb njohjes s\u00eb z\u00ebrit t\u00eb zhvilluar nga Mozilla DeepSpeech 0.6<\/a><\/noindex>, i cili implementon arkitektur\u00ebn e njohjes s\u00eb z\u00ebrit me t\u00eb nj\u00ebjtin em\u00ebr, t\u00eb propozuar<\/a><\/noindex> nga hulumtuesit e Baidu. Implementimi \u00ebsht\u00eb shkruar n\u00eb Python duke p\u00ebrdorur platform\u00ebn e m\u00ebsimit t\u00eb makinerive TensorFlow dhe distribuohet<\/a><\/noindex> n\u00ebn licenc\u00ebn e lir\u00eb MPL 2.0. Mb\u00ebshtet funksionimin n\u00eb Linux, Android, macOS dhe Windows. Performanca \u00ebsht\u00eb e mjaftueshme p\u00ebr t\u00eb p\u00ebrdorur motorin n\u00eb pajisjet LePotato, Raspberry Pi 3 dhe Raspberry Pi 4.<\/p>\n

N\u00eb paket\u00eb gjithashtu p\u00ebrfshihen ofruar<\/a><\/noindex> modele t\u00eb trajnuara, shembuj<\/a><\/noindex> skedar\u00eb audio dhe nj\u00eb mjet p\u00ebr njohjen p\u00ebrmes linj\u00ebs s\u00eb komand\u00ebs. P\u00ebr integrimin e funksionit t\u00eb njohjes s\u00eb z\u00ebrit n\u00eb programet e tua, jan\u00eb ofruar module t\u00eb gatshme p\u00ebr p\u00ebrdorim p\u00ebr Python, NodeJS, C++ dhe .NET (module t\u00eb p\u00ebrgatitura ve\u00e7mas nga zhvillues t\u00eb tret\u00eb p\u00ebr Rust<\/a><\/noindex> dhe Go<\/a><\/noindex>). Modeli i gatsh\u00ebm ofrohet vet\u00ebm p\u00ebr gjuh\u00ebn angleze, por p\u00ebr gjuh\u00eb t\u00eb tjera sipas dokumentacionit t\u00eb<\/a><\/noindex> udh\u00ebzues<\/a><\/noindex> mund t\u00eb trajnohet sistemi vet\u00eb, duke p\u00ebrdorur t\u00eb dh\u00ebna zanore<\/a><\/noindex>, t\u00eb mbledhura nga projekti Common Voice. <\/p>\n

DeepSpeech \u00ebsht\u00eb shum\u00eb m\u00eb i thjesht\u00eb se sistemet tradicionale dhe nj\u00ebkoh\u00ebsisht siguron nj\u00eb cil\u00ebsi m\u00eb t\u00eb lart\u00eb njohjeje n\u00eb prani t\u00eb zhurmave t\u00eb jashtme. N\u00eb zhvillim nuk p\u00ebrdoren modelet tradicionale akustike dhe koncepti i fonemave, por zbatohen nj\u00eb sistem m\u00ebsimi t\u00eb makinerive t\u00eb optimizuar mir\u00eb n\u00eb baz\u00eb t\u00eb nj\u00eb nervozi, i cili lejon shmangien e zhvillimit t\u00eb komponenteve t\u00eb ve\u00e7anta p\u00ebr modelimin e njohjes s\u00eb ndryshimeve t\u00eb ndryshme, si zhurma, jehona dhe karakteristikat e t\u00eb folurit. <\/p>\n

Ankthi i k\u00ebtij qasje \u00ebsht\u00eb se p\u00ebr t\u00eb arritur nj\u00eb njohje cil\u00ebsore dhe p\u00ebr t\u00eb trajnuar nervin e rrjetit, motorit DeepSpeech i nevojitet nj\u00eb sasi e madhe e t\u00eb dh\u00ebnave t\u00eb larmishme, t\u00eb diktuara n\u00eb kushte reale nga z\u00ebra t\u00eb ndrysh\u00ebm dhe n\u00eb prani t\u00eb zhurmave natyrore.
\nGrumbullimi i t\u00eb dh\u00ebnave t\u00eb tilla merret me projektin e krijuar n\u00eb Mozilla Common Voice<\/a><\/noindex>, duke ofruar nj\u00eb grup t\u00eb verifikuar t\u00eb dh\u00ebnash me 780 or\u00eb n\u00eb gjuh\u00ebn angleze<\/a><\/noindex>, 325 n\u00eb gjermanisht, 173 n\u00eb fr\u00ebngjisht dhe 27 or\u00eb n\u00eb rusisht. <\/p>\n

Q\u00ebllimi p\u00ebrfundimtar i projektit Common Voice \u00ebsht\u00eb akumulimi i 10,000 or\u00ebve me regjistrime t\u00eb shqiptimit t\u00eb frazave t\u00eb zakonshme t\u00eb njer\u00ebzve, gj\u00eb q\u00eb do t\u00eb lejoj\u00eb arritjen e nj\u00eb niveli t\u00eb pranuesh\u00ebm t\u00eb gabimeve n\u00eb njohje. N\u00eb gjendjen aktuale, pjes\u00ebmarr\u00ebsit e projektit tashm\u00eb kan\u00eb diktuar nj\u00eb total prej 4.3 mij\u00eb or\u00ebsh, nga t\u00eb cilat 3.5 mij\u00eb kan\u00eb kaluar verifikimin. Gjat\u00eb trajnimit t\u00eb modelit p\u00ebrfundimtar t\u00eb gjuh\u00ebs angleze p\u00ebr DeepSpeech jan\u00eb p\u00ebrdorur 3816 or\u00eb z\u00ebri, p\u00ebrve\u00e7 Common Voice, e cila mbulon t\u00eb dh\u00ebna nga projektet LibriSpeech, Fisher dhe Switchboard, si dhe p\u00ebrfshin rreth 1700 or\u00eb regjistrimesh t\u00eb transkripuara nga emisionet radio.<\/p>\n

P\u00ebr p\u00ebrdorimin e modelit t\u00eb gatsh\u00ebm p\u00ebr shkarkim t\u00eb gjuh\u00ebs angleze, niveli i gabimeve t\u00eb njohjes n\u00eb DeepSpeech \u00ebsht\u00eb 7.5% sipas vler\u00ebsimit t\u00eb grupit t\u00eb testit LibriSpeech<\/a><\/noindex>. P\u00ebr krahasim, niveli i gabimeve n\u00eb njohjen e njeriut vler\u00ebsohet<\/a><\/noindex> \u00ebsht\u00eb 5.83%.<\/p>\n

DeepSpeech p\u00ebrb\u00ebhet nga dy n\u00ebn-sisteme - modeli akustik dhe dekoduesi. Modeli akustik p\u00ebrdor metoda t\u00eb m\u00ebsimit t\u00eb thell\u00eb t\u00eb makin\u00ebs p\u00ebr t\u00eb llogaritur probabilitetin e pranis\u00eb s\u00eb simboleve t\u00eb caktuara n\u00eb z\u00ebrin e dh\u00ebn\u00eb si input. Dekoduesi aplikon nj\u00eb algorit\u00ebm t\u00eb k\u00ebrkimit me rreze p\u00ebr t\u00eb transformuar t\u00eb dh\u00ebnat mbi probabilitetin e simboleve n\u00eb nj\u00eb p\u00ebrfaq\u00ebsim tekstual.<\/p>\n

Ndryshimet kryesore risit\u00eb<\/a><\/noindex> DeepSpeech 0.6 (dega 0.6 nuk \u00ebsht\u00eb e p\u00ebrputhshme me versione e kaluara dhe k\u00ebrkon p\u00ebrdit\u00ebsimin e kodit dhe modeleve):<\/p>\n