I-Google ikhiphe i-V2 ye-Lyra, i-codec yomthombo ovulekile we-bitrate ephansi

I-Lyra i-codec yomsindo ye-Google

I-Google ikhiphe inguqulo yesibili ye-Lyra, i-codec yayo yekhwalithi ephezulu, ene-bitrate ephansi eyenza ukuxhumana ngezwi kutholakale ngisho nakumanethiwekhi ahamba kancane.

Muva nje I-Google yembulwe ngeposi lebhulogi, ikhipha inguqulo yesibili yekhodekhi yakho yomsindo "Lyra-V2", esebenzisa amasu okufunda omshini ukuze afinyelele izinga eliphezulu lezwi lapho usebenzisa iziteshi zokuxhumana ezihamba kancane kakhulu.

Uhlobo olusha yethula inguquko ekwakhiweni okusha kwenethiwekhi ye-neural, ukusekelwa kwezinkundla ezengeziwe, isilawuli se-bitrate esithuthukisiwe, ukuthuthukiswa kokusebenza, nekhwalithi yomsindo ephezulu.

Manje sikhulula i-Lyra V2, enesakhiwo esisha esijabulela ukwesekwa kwenkundla ebanzi, esihlinzeka ngamakhono anyukayo e-bitrate, ukusebenza okungcono, nomsindo wekhwalithi ephezulu. Ngalokhu kukhululwa, sibheke ngabomvu ukuqhubeka nokuvela nomphakathi futhi, ngobuhlakani bakho obuhlangene, sibone izinhlelo zokusebenza ezintsha ezakhiwayo kanye nezikhombisi-ndlela ezintsha ezivelayo.

Mayelana noLyra

Mayelana nekhwalithi yedatha yezwi edluliswa ngesivinini esiphansi, I-Lyra iphakeme kakhulu kunamakhodekhi endabuko abasebenzisa izindlela zokucubungula isignali yedijithali. Ukuze kuzuzwe ukudluliswa kwezwi kwekhwalithi ephezulu ngaphansi kwezimo zenani elilinganiselwe lolwazi oludlulisiwe, ngaphezu kokucindezelwa komsindo okujwayelekile nezindlela zokuguqula isignali, ULyra usebenzisa imodeli yezwi esekelwe ohlelweni lokufunda lomshini okukuvumela ukuthi udale kabusha ulwazi olulahlekile. ngokusekelwe ezicini zenkulumo ezijwayelekile.

Ikhodekhi ihlanganisa isifaki khodi nesikhikhoda. I-algorithm yesifaki khodi ikhipha amapharamitha edatha yezwi njalo ngama-milliseconds angu-20, iwacindezele futhi iwadlulisele kumamukeli kunethiwekhi enesilinganiso esincane esingu-3,2 kbps ukuya ku-9,2 kbps.

Ohlangothini lomamukeli, idekhoda isebenzisa imodeli ekhiqizayo ukuze idale kabusha isignali yenkulumo yoqobo ngokusekelwe kumapharamitha omsindo odluliswayo, okuhlanganisa ama-spectrogram we-logarithmic choki acabangela izici zamandla enkulumo kumabanga ahlukene wefrikhwensi. futhi alungiselelwa kucatshangwa ngombono womuntu .

Yini entsha ku-Lyra V2?

I-Lyra V2 isebenzisa imodeli entsha yokukhiqiza esekelwe kunethiwekhi ye-SoundStream neural, enezidingo eziphansi zokubala, ezivumela ukuqoshwa kwesikhathi sangempela ngisho nasezinhlelweni zamandla aphansi.

Imodeli esetshenziselwa ukukhiqiza umsindo iqeqeshwe kusetshenziswa izinkulungwane ezimbalwa zamahora wokuqoshwa kwezwi ngezilimi ezingaphezu kuka-90 (I-TensorFlow Lite isetshenziselwa ukuqalisa imodeli). Ukusebenza kokuqaliswa okuhlongozwayo kwanele ukufaka ikhodi futhi kuqondwe izwi kuma-smartphones ebanga lentengo eliphansi kakhulu.

Ngaphezu kokusebenzisa imodeli ehlukile yokukhiqiza, inguqulo entsha iphinde ivelele ekufakweni kwezixhumanisi ne-RVQ quantifier (I-Residual Vector Quantizer) ekwakhiweni kwekhodekhi, eyenziwa ngasohlangothini lomthumeli ngaphambi kokudluliswa kwedatha, nasohlangothini lomamukeli ngemva kokwamukela idatha.

I-quantizer iguqula imingcele enikezwe i-codec ibe amasethi amaphakethe, ibhala ngekhodi ulwazi oluhlobene nesilinganiso sebhithi esikhethiwe. Ukuqinisekisa amaleveli ekhwalithi ahlukene, ama-quantizer ahlinzekwa ngama-bitrate amathathu (3,2kbps, 6kbps, kanye no-9,2kbps), uma i-bitrate iphezulu, ikhwalithi engcono kakhulu, kodwa iba phezulu izidingo zomkhawulokudonsa.

i-architecture entsha yehlise ukubambezeleka kokudluliselwa kwesignali ukusuka kumamillisecond angu-100 kuya ku-20 millisecond. Uma kuqhathaniswa, i-Opus codec ye-WebRTC ibonise ukubambezeleka okungu-26,5 ms, 46,5 ms, no-66,5 ms ngamanani amancane ahloliwe. Ukusebenza kwesifaki khodi nesikhikhikhoda nakho kukhuphuke kakhulu: Uma kuqhathaniswa nenguqulo yangaphambilini, kukhona ukusheshisa okufika ezikhathini ezi-5. Isibonelo, ku-smartphone ye-Pixel 6 Pro, i-codec entsha ifaka ikhodi futhi iqophe isampula engu-20ms ngo-0,57ms, eshesha izikhathi ezingu-35 kunalokho okudingekayo ekusakazeni kwesikhathi sangempela.

Ngaphezu kokusebenza, siphinde sakwazi ukuthuthukisa ikhwalithi yokubuyisela umsindo: ngokwesilinganiso se-MUSHRA, ikhwalithi yenkulumo ngamabhithi angu-3,2 kbps, 6 kbps no-9,2 kbps uma usebenzisa i-codec ye-Lyra V2 ihambisana nezilinganiso zebhithi ezingu-10 kbps, 13 kbps no-14 kbps uma usebenzisa i-Opus codec.

Okokugcina uma unentshisekelo yokwazi kabanzi ngakho, ungabheka imininingwane ku- isixhumanisi esilandelayo.


Shiya umbono wakho

Ikheli lakho le ngeke ishicilelwe. Ezidingekayo ibhalwe nge *

*

*

  1. Ubhekele imininingwane: Miguel Ángel Gatón
  2. Inhloso yedatha: Lawula Ugaxekile, ukuphathwa kwamazwana.
  3. Ukusemthethweni: Imvume yakho
  4. Ukuxhumana kwemininingwane: Imininingwane ngeke idluliselwe kubantu besithathu ngaphandle kwesibopho esisemthethweni.
  5. Isitoreji sedatha: Idatabase ebanjwe yi-Occentus Networks (EU)
  6. Amalungelo: Nganoma yisiphi isikhathi ungakhawulela, uthole futhi ususe imininingwane yakho.