Spole\u010dnost Google p\u0159edstavila audiokodek Lyra V2, kter\u00fd vyu\u017e\u00edv\u00e1 metody strojov\u00e9ho u\u010den\u00ed k dosa\u017een\u00ed maxim\u00e1ln\u00ed kvality p\u0159enosu \u0159e\u010di p\u0159i pou\u017eit\u00ed velmi pomal\u00fdch komunika\u010dn\u00edch kan\u00e1l\u016f. Nov\u00e1 verze se vyzna\u010duje p\u0159echodem na novou architekturu neuronov\u00e9 s\u00edt\u011b, podporou dal\u0161\u00edch platforem, roz\u0161\u00ed\u0159en\u00fdmi mo\u017enostmi \u0159\u00edzen\u00ed bitrate, zv\u00fd\u0161en\u00edm v\u00fdkonu a dosa\u017een\u00edm vy\u0161\u0161\u00ed kvality zvuku. Referen\u010dn\u00ed implementace k\u00f3du je naps\u00e1na v C++ a \u0161\u00ed\u0159\u00ed se pod licenc\u00ed Apache 2.0. <\/p>\n
Z hlediska kvality p\u0159en\u00e1\u0161en\u00fdch hlasov\u00fdch dat p\u0159i n\u00edzk\u00fdch rychlostech Lyra v\u00fdrazn\u011b p\u0159ekon\u00e1v\u00e1 tradi\u010dn\u00ed kodeky, kter\u00e9 pou\u017e\u00edvaj\u00ed metody digit\u00e1ln\u00edho zpracov\u00e1n\u00ed sign\u00e1l\u016f. Pro dosa\u017een\u00ed vysok\u00e9 kvality p\u0159enosu hlasu za podm\u00ednek omezen\u00e9ho objemu p\u0159en\u00e1\u0161en\u00fdch informac\u00ed, krom\u011b b\u011b\u017en\u00fdch metod komprese zvuku a transformace sign\u00e1l\u016f, Lyra pou\u017e\u00edv\u00e1 hlasov\u00fd model zalo\u017een\u00fd na syst\u00e9mu strojov\u00e9ho u\u010den\u00ed, kter\u00fd umo\u017e\u0148uje rekonstrukci chyb\u011bj\u00edc\u00edch informac\u00ed na z\u00e1klad\u011b typick\u00fdch charakteristik \u0159e\u010di. <\/p>\n
Kodek zahrnuje kod\u00e9r a dekod\u00e9r. Algoritmus fungov\u00e1n\u00ed kod\u00e9ru spo\u010d\u00edv\u00e1 v extrakci parametr\u016f hlasov\u00fdch dat ka\u017ed\u00fdch 20 milisekund, jejich kompresi a p\u0159ed\u00e1n\u00ed p\u0159\u00edjemci p\u0159es s\u00ed\u0165 s bitrate od 3,2 kbps do 9,2 kbps. Na stran\u011b p\u0159\u00edjemce dekod\u00e9r pou\u017e\u00edv\u00e1 generativn\u00ed model k rekonstrukci p\u016fvodn\u00edho \u0159e\u010dov\u00e9ho sign\u00e1lu na z\u00e1klad\u011b p\u0159enesen\u00fdch zvukov\u00fdch parametr\u016f, kter\u00e9 zahrnuj\u00ed logaritmick\u00e9 mel-spektrogramy zohled\u0148uj\u00edc\u00ed charakteristiky energie \u0159e\u010di v r\u016fzn\u00fdch frekven\u010dn\u00edch p\u00e1smech a p\u0159ipraven\u00e9 s ohledem na model lidsk\u00e9ho sluchov\u00e9ho vn\u00edm\u00e1n\u00ed. <\/p>\n
V Lyra V2 byla vyu\u017eita nov\u00e1 generativn\u00ed model na b\u00e1zi konvolu\u010dn\u00ed neuronov\u00e9 s\u00edt\u011b SoundStream, kter\u00e1 se vyzna\u010duje n\u00edzk\u00fdmi n\u00e1roky na v\u00fdpo\u010detn\u00ed zdroje, co\u017e umo\u017e\u0148uje dek\u00f3dov\u00e1n\u00ed v re\u00e1ln\u00e9m \u010dase i na slab\u0161\u00edch syst\u00e9mech. Model pou\u017eit\u00fd pro generaci zvuku byl vy\u0161kolen pomoc\u00ed n\u011bkolika tis\u00edc hodin z\u00e1znam\u016f hlas\u016f ve v\u00edce ne\u017e 90 jazyc\u00edch. Pro prov\u00e1d\u011bn\u00ed modelu se vyu\u017e\u00edv\u00e1 TensorFlow Lite. V\u00fdkon navr\u017een\u00e9 implementace je dostate\u010dn\u00fd pro k\u00f3dov\u00e1n\u00ed a dek\u00f3dov\u00e1n\u00ed \u0159e\u010di na smartphonech ni\u017e\u0161\u00ed cenov\u00e9 kategorie. <\/p>\n
Krom\u011b vyu\u017eit\u00ed odli\u0161n\u00e9ho generativn\u00edho modelu je nov\u00e1 verze tak\u00e9 zaj\u00edmav\u00e1 zahrnut\u00edm v architektu\u0159e kodeku \u010dl\u00e1nk\u016f s kvantiz\u00e1torem RVQ (Residual Vector Quantizer), kter\u00fd se prov\u00e1d\u00ed na stran\u011b odes\u00edlatele p\u0159ed p\u0159enosem dat a na stran\u011b p\u0159\u00edjemce po p\u0159ijet\u00ed dat. Kvantiz\u00e1tor p\u0159ev\u00e1d\u00ed parametry vyd\u00e1van\u00e9 kodekem na soubory paket\u016f, k\u00f3dov\u00e1n\u00edm informac\u00ed podle zvolen\u00e9ho bitratu. Pro zaji\u0161t\u011bn\u00ed r\u016fzn\u00e9 \u00farovn\u011b kvality jsou k dispozici kvantiz\u00e1tory pro t\u0159i bitraty (3,2 kps, 6 kbps a 9,2 kbps), \u010d\u00edm vy\u0161\u0161\u00ed bitrate, t\u00edm lep\u0161\u00ed kvalita, ale vy\u0161\u0161\u00ed po\u017eadavky na \u0161\u00ed\u0159ku p\u00e1sma.
<\/center> <\/p>\n
Nov\u00e1 architektura umo\u017enila sn\u00ed\u017eit zpo\u017ed\u011bn\u00ed p\u0159enosu sign\u00e1lu z 100 na 20 milisekund. Pro srovn\u00e1n\u00ed, kodek Opus pro WebRTC vyk\u00e1zal p\u0159i testovan\u00fdch bitrat\u00edch zpo\u017ed\u011bn\u00ed 26,5 ms, 46,5 ms a 66,5 ms. V\u00fdrazn\u011b se tak\u00e9 zv\u00fd\u0161ila v\u00fdkonnost kod\u00e9ru a dekod\u00e9ru \u2014 ve srovn\u00e1n\u00ed s p\u0159edchoz\u00ed verz\u00ed bylo zaznamen\u00e1no zrychlen\u00ed a\u017e 5kr\u00e1t. Nap\u0159\u00edklad na smartphonu Pixel 6 Pro nov\u00fd kodek prov\u00e1d\u00ed k\u00f3dov\u00e1n\u00ed a dek\u00f3dov\u00e1n\u00ed 20milisekundov\u00e9ho vzorku za 0,57 ms, co\u017e je 35kr\u00e1t rychlej\u0161\u00ed, ne\u017e je pot\u0159eba pro p\u0159enos v re\u00e1ln\u00e9m \u010dase. <\/p>\n
Krom\u011b zv\u00fd\u0161en\u00ed v\u00fdkonu se poda\u0159ilo dos\u00e1hnout i zlep\u0161en\u00ed kvality obnovy zvuku \u2014 na \u0161k\u00e1le MUSHRA odpov\u00edd\u00e1 kvalita \u0159e\u010di p\u0159i bitratech 3,2 kbps, 6 kbps a 9,2 kbps p\u0159i pou\u017eit\u00ed kodeku Lyra V2 bitrate\u016fm 10 kbps, 13 kbps a 14 kbps p\u0159i pou\u017eit\u00ed kodeku Opus. \n Zdroj: 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