Facebook imasindikiza ma codec omvera a EnCodec pogwiritsa ntchito kuphunzira pamakina

Meta / Facebook (yoletsedwa ku Russian Federation) inayambitsa codec yatsopano ya audio, EnCodec, yomwe imagwiritsa ntchito njira zophunzirira makina kuti ziwonjezere chiwerengero cha kuponderezana popanda kutaya khalidwe. Ma codec amatha kugwiritsidwa ntchito potulutsa mawu munthawi yeniyeni komanso posungira kuti musunge mafayilo. Kukhazikitsa kwa EnCodec kumalembedwa mu Python pogwiritsa ntchito PyTorch framework ndipo ndi chilolezo pansi pa CC BY-NC 4.0 (Creative Commons Attribution-NonCommercial) laisensi yongogwiritsa ntchito osati malonda okha.

Mitundu iwiri yopangidwa kale imaperekedwa kuti itsitsidwe:

  • Mtundu woyambitsa wogwiritsa ntchito 24 kHz sampling rate, wothandizira mawu omvera okha, komanso wophunzitsidwa pamitundu yosiyanasiyana yamawu (yoyenera kuyika mawu). Mtunduwu ungagwiritsidwe ntchito kuyika deta yamawu kuti itumizidwe pamitengo pang'ono ya 1.5, 3, 6, 12 ndi 24 kbps.
  • Mtundu wopanda chifukwa chogwiritsa ntchito zitsanzo za 48 kHz, zothandizira ma audio a stereo komanso ophunzitsidwa nyimbo zokha. Mtunduwu umathandizira ma bitrate a 3, 6, 12 ndi 24 kbps.

Pachitsanzo chilichonse, chinenero chowonjezera chakonzedwa, chomwe chimakulolani kuti mukwaniritse kuwonjezeka kwakukulu kwa chiwerengero cha kuponderezedwa (mpaka 40%) popanda kutaya khalidwe. Mosiyana ndi mapulojekiti omwe adapangidwa kale pogwiritsa ntchito njira zophunzirira zamakina zophatikizira mawu, EnCodec itha kugwiritsidwa ntchito osati kungoyika mawu, komanso kuphatikizika kwanyimbo ndi zitsanzo za 48 kHz, zogwirizana ndi kuchuluka kwa ma CD omvera. Malinga ndi omwe akupanga codec yatsopano, potumiza ndi bitrate ya 64 kbps poyerekeza ndi mtundu wa MP3, adatha kukulitsa kuchuluka kwa kupsinjika kwamawu pafupifupi kakhumi ndikusunga mulingo womwewo (mwachitsanzo, pogwiritsa ntchito MP3, bandiwifi ya 64 kbps ndiyofunika, kuti kufalitsa ndi mtundu womwewo wa EnCodec ndikokwanira 6 kbps).

Zomangamanga za codec zimamangidwa pa neural network yokhala ndi "transformer" yomanga ndipo imachokera pamalumikizidwe anayi: encoder, quantizer, decoder ndi discriminator. Encoder imachotsa magawo a data ya mawu ndikusintha mtsinje wodzaza kukhala wocheperako. Quantizer (RVQ, Residual Vector Quantizer) imasintha zomwe zimatuluka ndi encoder kukhala mapaketi, kukakamiza zambiri kutengera bitrate yosankhidwa. Kutulutsa kwa quantizer ndi chiwonetsero choponderezedwa cha data, choyenera kutumizidwa pa netiweki kapena kusungidwa ku disk.

The decoder decode kuimilira kokanikizidwa kwa deta ndikumanganso mafunde apachiyambi. Wosankhana amapititsa patsogolo ubwino wa zitsanzo zopangidwa, poganizira chitsanzo cha kulingalira kwa anthu. Mosasamala kanthu za mtundu ndi bitrate, mitundu yomwe imagwiritsidwa ntchito pokopera ndi kusindikiza imasiyanitsidwa ndi zofunikira zocheperako (kuwerengera kofunikira pakugwirira ntchito zenizeni kumachitika pakatikati pa CPU).

Facebook imasindikiza ma codec omvera a EnCodec pogwiritsa ntchito kuphunzira pamakina


Source: opennet.ru

Kuwonjezera ndemanga