Kompania Google prezantoi kodekun audio Lyra V2, i cili p\u00ebrdor metoda t\u00eb m\u00ebsimit t\u00eb makineris\u00eb p\u00ebr t\u00eb arritur cil\u00ebsin\u00eb maksimale t\u00eb transmetimit t\u00eb z\u00ebrit duke p\u00ebrdorur kanale t\u00eb ngadalta komunikimi. Versioni i ri karakterizohet nga kalimi n\u00eb nj\u00eb arkitektur\u00eb t\u00eb re t\u00eb rrjetit nervor, mb\u00ebshtetje p\u00ebr platforma shtes\u00eb, mund\u00ebsi t\u00eb zgjeruara p\u00ebr menaxhimin e bitrate, rritje t\u00eb performanc\u00ebs dhe arritje t\u00eb nj\u00eb cil\u00ebsie m\u00eb t\u00eb lart\u00eb t\u00eb tingullit. Implementimi referenc\u00eb i kodit \u00ebsht\u00eb i shkruar n\u00eb C++ dhe shp\u00ebrndahet n\u00ebn licenc\u00ebn Apache 2.0. <\/p>\n
Sa i p\u00ebrket cil\u00ebsis\u00eb s\u00eb t\u00eb dh\u00ebnave t\u00eb z\u00ebrit t\u00eb transmetuara n\u00eb shpejt\u00ebsi t\u00eb ul\u00ebta, Lyra tejkalon ndjesh\u00ebm kodek\u00ebt tradicional\u00eb q\u00eb p\u00ebrdorin metoda t\u00eb p\u00ebrpunimit dixhital t\u00eb sinjalit. P\u00ebr t\u00eb arritur cil\u00ebsin\u00eb e lart\u00eb t\u00eb transmetimit t\u00eb z\u00ebrit n\u00eb kushte t\u00eb informacionit t\u00eb kufizuar, p\u00ebrve\u00e7 metodave t\u00eb zakonshme t\u00eb kompresimit t\u00eb z\u00ebrit dhe shnd\u00ebrrimit t\u00eb sinjaleve, n\u00eb Lyra aplikohet nj\u00eb model z\u00ebri me baz\u00eb n\u00eb sistemin e m\u00ebsimit t\u00eb makineris\u00eb, i cili lejon rip\u00ebrtritjen e informacionit t\u00eb munguar n\u00eb baz\u00eb t\u00eb karakteristikave tipike t\u00eb z\u00ebrit. <\/p>\n
Kodeku p\u00ebrfshin nj\u00eb kodues dhe nj\u00eb dekodues. Algoritmi i pun\u00ebs s\u00eb koduesit konsiston n\u00eb nxjerrjen e parametrave t\u00eb t\u00eb dh\u00ebnave t\u00eb z\u00ebrit \u00e7do 20 milisekonda, kompresimin e tyre dhe d\u00ebrgimin e tyre te marr\u00ebsi p\u00ebrmes rrjetit me nj\u00eb bit rate nga 3.2kbps deri n\u00eb 9.2kbps. N\u00eb an\u00ebn e marr\u00ebsit, dekoduesi p\u00ebrdor nj\u00eb model gjenerativ p\u00ebr t\u00eb riprodhuar sinjalin origjinal t\u00eb z\u00ebrit bazuar n\u00eb parametrat e z\u00ebrit t\u00eb d\u00ebrguar, t\u00eb cil\u00ebt p\u00ebrfshijn\u00eb spektrat logaritmik\u00eb mel, q\u00eb marrin parasysh karakteristikat e energjis\u00eb s\u00eb z\u00ebrit n\u00eb frekuenca t\u00eb ndryshme dhe jan\u00eb p\u00ebrgatitur n\u00eb p\u00ebrputhje me modelin e perceptimit t\u00eb d\u00ebgjimit t\u00eb njeriut. <\/p>\n
N\u00eb Lyra V2 \u00ebsht\u00eb p\u00ebrdorur nj\u00eb model i ri gjenerativ i bazuar n\u00eb rrjetin e neuroneve konvencionale SoundStream, i cili dallon p\u00ebr k\u00ebrkesat e ul\u00ebta p\u00ebr burime kompjuterike, duke mund\u00ebsuar k\u00ebshtu dekodimin n\u00eb koh\u00eb reale edhe n\u00eb sisteme me performanc\u00eb t\u00eb ul\u00ebt. Modeli i p\u00ebrdorur p\u00ebr gjenerimin e z\u00ebrit \u00ebsht\u00eb trajnuar duke p\u00ebrdorur disa mij\u00ebra or\u00eb regjistrimesh t\u00eb z\u00ebra n\u00eb m\u00eb shum\u00eb se 90 gjuh\u00eb. P\u00ebr realizimin e modelit p\u00ebrdoret TensorFlow Lite. Performanca e zgjidhjes s\u00eb propozuar \u00ebsht\u00eb e mjaftueshme p\u00ebr kodimin dhe dekodimin e z\u00ebrit n\u00eb smartphone t\u00eb gam\u00ebs s\u00eb ul\u00ebt. <\/p>\n
P\u00ebrve\u00e7 p\u00ebrdorimit t\u00eb nj\u00eb modeli tjet\u00ebr gjenerativ, versioni i ri \u00ebsht\u00eb ve\u00e7an\u00ebrisht i njohur p\u00ebr p\u00ebrfshirjen n\u00eb arkitektur\u00ebn e kodek\u00ebve t\u00eb lidhjeve me kvantizuesin RVQ (Residual Vector Quantizer), i cili kryhet n\u00eb an\u00ebn e d\u00ebrguesit para d\u00ebrgimit t\u00eb t\u00eb dh\u00ebnave dhe n\u00eb an\u00ebn e pranuesit pas marrjes s\u00eb t\u00eb dh\u00ebnave. Kvantizuesi kthen parametrat e dh\u00ebn\u00eb nga kodeku n\u00eb grupe paketeshe, duke koduar informacionin n\u00eb lidhje me bitratin e zgjedhur. P\u00ebr t\u00eb siguruar nivele t\u00eb ndryshme cil\u00ebsie, jan\u00eb parashikuar kvantizues p\u00ebr tre bitrate (3.2 kps, 6 kbps dhe 9.2 kbps); sa m\u00eb i lart\u00eb t\u00eb jet\u00eb bitrate, aq m\u00eb e mir\u00eb \u00ebsht\u00eb cil\u00ebsia, por k\u00ebrkesa p\u00ebr band\u00eb \u00ebsht\u00eb m\u00eb e lart\u00eb.
<\/center> <\/p>\n
Arqitektura e re e mund\u00ebsoi uljen e vonesave t\u00eb transmetimit t\u00eb sinjalit nga 100 n\u00eb 20 milisekonda. P\u00ebr krahasim, kodeku Opus p\u00ebr WebRTC tregoi vonesa n\u00eb bitrate t\u00eb testuara prej 26.5ms, 46.5ms dhe 66.5ms. Gjithashtu, performanca e enkoderit dhe dekoderit \u00ebsht\u00eb rritur ndjesh\u00ebm \u2014 krahasuar me versionin e kaluar, \u00ebsht\u00eb v\u00ebrejtur nj\u00eb p\u00ebrshpejtim deri n\u00eb 5 her\u00eb. P\u00ebr shembull, n\u00eb smartphone-in Pixel 6 Pro, kodeku i ri realizon kodimin dhe dekodimin e mostr\u00ebs 20-milisekond\u00ebshe p\u00ebr 0.57 ms, q\u00eb \u00ebsht\u00eb 35 her\u00eb m\u00eb shpejt se sa nevojitet p\u00ebr transmetim n\u00eb koh\u00eb reale. <\/p>\n
P\u00ebrve\u00e7 performanc\u00ebs, arriti gjithashtu nj\u00eb p\u00ebrmir\u00ebsim n\u00eb cil\u00ebsin\u00eb e rikuperimit t\u00eb tingullit \u2014 n\u00eb shkall\u00ebn MUSHRA, cil\u00ebsia e z\u00ebrit n\u00eb bitrate 3.2 kbps, 6 kbps dhe 9.2 kbps duke p\u00ebrdorur kodekun Lyra V2 korrespondon me bitrate 10 kbps, 13 kbps dhe 14 kbps duke p\u00ebrdorur kodekun Opus. \n Burimi: opennet.ru<\/a> <\/p>","protected":false,"gt_translate_keys":[{"key":"rendered","format":"html"}]},"excerpt":{"rendered":"
![\"Google](\"\/wp-content\/uploads\/2022\/10\/54943fd2b2c5c2d008c8c35e462f4a4b.jpg\")
<\/center> <\/p>\n