Google ewebata codec ọdịyo ọhụrụ, Lyra, ahaziri iji nweta ogo olu kachasị ọbụlagodi mgbe ị na-eji ọwa nzikọrịta ozi dị nwayọ. Edere koodu mmejuputa Lyra na C++ wee mepee n'okpuru ikike Apache 2.0, mana n'ime ihe ndabere achọrọ maka ịrụ ọrụ enwere ọbá akwụkwọ libsparse_inference.so nwere mmejuputa kernel maka mgbakọ na mwepụ. A na-achọpụta na ọbá akwụkwọ nke nwe ya bụ nwa oge - n'ọdịnihu Google na-ekwe nkwa ịmepụta onye na-anọchi anya ma nye nkwado maka ikpo okwu dị iche iche.
N'ihe gbasara ogo data olu a na-ebufe na obere ọsọ, Lyra dị elu karịa codecs ọdịnala na-eji ụzọ nhazi akara dijitalụ. Iji nweta mgbasa ozi olu dị elu na ọnọdụ nke obere ozi ebugharị, na mgbakwunye na usoro ọdịnala nke mkpakọ ọdịyo na ngbanwe mgbaàmà, Lyra na-eji usoro okwu dabere na sistemụ mmụta igwe, nke na-enye gị ohere ịmegharị ozi na-efu efu dabere na ya. njirimara okwu ahụkarị. A zụrụ ihe nlereanya eji emepụta ụda a zụrụ site na iji ọtụtụ puku awa nke ndekọ olu n'ihe karịrị asụsụ 70.
Codec ahụ gụnyere ihe ngbanwe na ihe ngbanwe. Algọridim nke ihe ngbanwe ahụ na-agbada ruo iwepụta paramita data olu kwa millisecond 40, na-akpakọ ha, ma na-ebufe ya na onye nnata na netwọkụ. Ọwa nzikọrịta ozi nwere ọsọ nke kilobits 3 kwa sekọnd zuru oke maka nnyefe data. Akụkụ ọdịyo ewepụtara gụnyere spectrograms logarithmic mel nke na-eburu n'uche njirimara ike nke okwu n'ụdị ọnụọgụ dị iche iche ma kwadebere na-eburu n'uche ihe atụ nke nghọta mmadụ.
Ihe ngbanwe ahụ na-eji ụdị nrụpụta nke, dabere na paramita ọdịyo ebugharị, na-emegharị mgbaama okwu. Iji belata mgbagwoju anya nke ngụkọ, a na-eji ihe nlereanya dị fechaa dabere na netwọkụ akwara na-emegharị ugboro ugboro, nke bụ ụdị dị iche iche nke ụdị okwu WaveRNN, nke na-eji obere nlele nlele dị ala, ma na-emepụta ọtụtụ akara n'otu n'otu n'ogo dị iche iche. A na-atụkwasịkwa akara ngosi ndị a kapụta ka ha mepụta otu mgbaama mmepụta dabara na ọnụego nlele akọwapụtara.
A na-ejikwa ntuziaka nhazi pụrụ iche dị na nhazi 64-bit ARM maka osooso. N'ihi ya, n'agbanyeghị iji igwe mmụta, enwere ike iji Lyra codec maka ntinye okwu na ngbanwe na ekwentị n'etiti, na-egosi nkwụsịtụ nnyefe nke 90 milliseconds.
isi: opennet.ru