Google e hlahisitse codec e ncha ea molumo, Lyra, e ntlafalitsoeng ho fihlela boleng ba lentsoe le ha o sebelisa mecha ea puisano e liehang haholo. Khoutu ea ts'ebetsong ea Lyra e ngotsoe ka C++ 'me e butsoe tlas'a laesense ea Apache 2.0, empa har'a litšepiso tse hlokahalang bakeng sa ts'ebetso ho na le laebrari ea thepa libsparse_inference.kahoo ka ts'ebetsong ea kernel bakeng sa lipalo tsa lipalo. Hoa hlokomeleha hore laebrari ea thepa ke ea nakoana - nakong e tlang Google e tšepisa ho hlahisa sebaka se bulehileng le ho fana ka tšehetso bakeng sa liforomo tse fapaneng.
Mabapi le boleng ba data ea lentsoe e fetisitsoeng ka lebelo le tlaase, Lyra e phahame haholo ho li-codec tsa setso tse sebelisang mekhoa ea ho sebetsana le matšoao a digital. Ho finyella phetiso ea lentsoe la boleng bo phahameng maemong a palo e fokolang ea boitsebiso bo fetisitsoeng, ho phaella mekhoeng e tloaelehileng ea ho hatella molumo le ho fetola pontšo, Lyra o sebelisa mokhoa oa ho bua o thehiloeng ho mokhoa oa ho ithuta oa mochine, o u lumellang ho tsosolosa boitsebiso bo sieo bo thehiloeng ho litšobotsi tse tloaelehileng tsa puo. Mokhoa o sebelisitsoeng ho hlahisa molumo o ile oa koetlisoa ka lihora tse likete tse 'maloa tsa ho rekota mantsoe ka lipuo tse fetang 70.
Codec e kenyelletsa encoder le decoder. Algorithm ea "encoder" e tsoela pele ho ntša liparamente tsa data ea lentsoe ka li-milliseconds tse ling le tse ling tse 40, ho li hatella, le ho li fetisetsa ho moamoheli marang-rang. Mocha oa puisano o nang le lebelo la lik'hilograma tse 3 ka motsotsoana o lekane bakeng sa ho fetisetsa data. Mekhahlelo ea molumo e ntšitsoeng e kenyelletsa li-spectrogram tsa logarithmic mel tse nkang matla a matla a puo ka mekhahlelo e fapaneng ea maqhubu 'me li lokiselitsoe ho ela hloko mohlala oa kutloisiso ea kutlo ea motho.
Decoder e sebelisa mofuta oa tlhahiso eo, ho ipapisitsoe le liparamente tsa molumo tse fetisoang, e bopang lets'oao la puo. E le ho fokotsa ho rarahana ha lipalo, ho ile ha sebelisoa mohlala o bobebe o thehiloeng ho marang-rang a tloaelehileng a neural, e leng mofuta o fapaneng oa WaveRNN oa synthesis ea puo, o sebelisang maqhubu a tlaase a sampling, empa o hlahisa lipontšo tse 'maloa ka ho tšoana ka mefuta e fapaneng ea maqhubu. Lipontšo tse hlahisoang li beoa holimo ho hlahisa lets'oao le le leng la tlhahiso le tsamaellanang le sekhahla se boletsoeng sa sampole.
Litaelo tse khethehileng tsa processor tse fumanehang ho li-processor tsa 64-bit ARM le tsona li sebelisetsoa ho potlakisa. Ka lebaka leo, ho sa tsotellehe ts'ebeliso ea ho ithuta ka mochine, Lyra codec e ka sebelisoa bakeng sa khouto ea puo ea nako ea sebele le ho khetholla li-smartphone tsa mahareng, ho bontša latency ea phetisetso ea matšoao ea 90 milliseconds.
Source: opennet.ru