Facebook tshaj tawm EnCodec audio codec siv tshuab kev kawm

Meta / Facebook (txwv tsis pub nyob rau hauv Lavxias Federation) tau tshaj tawm lub suab codec tshiab, EnCodec, uas siv cov kev kawm tshuab kom nce qhov sib piv tsis poob zoo. Lub codec tuaj yeem siv tau ob qho tib si rau streaming suab hauv lub sijhawm tiag tiag thiab rau encoding rau tom qab txuag hauv cov ntaub ntawv. Kev siv EnCodec yog sau rau hauv Python siv PyTorch lub moj khaum thiab tau tso cai raws li CC BY-NC 4.0 (Creative Commons Attribution-NonCommercial) daim ntawv tso cai rau kev siv tsis yog lag luam nkaus xwb.

Muaj ob hom kev npaj ua tiav rau kev rub tawm:

  • Tus qauv tsim qauv siv 24 kHz sampling tus nqi, txhawb nqa lub suab monophonic nkaus xwb, thiab tau kawm txog ntau hom suab ntaub ntawv (tsim rau kev hais lus coding). Tus qauv tuaj yeem siv los ntim cov ntaub ntawv audio rau kev sib kis ntawm tus nqi me ntsis ntawm 1.5, 3, 6, 12 thiab 24 kbps.
  • Tus qauv uas tsis yog vim li cas siv tus qauv piv txwv ntawm 48 kHz, txhawb nqa lub tso suab thiab kev cob qhia nkaus xwb ntawm cov suab paj nruag. Tus qauv txhawb bitrates ntawm 3, 6, 12 thiab 24 kbps.

Rau txhua tus qauv, ib qho qauv lus ntxiv tau raug npaj, uas tso cai rau koj kom ua tiav qhov kev sib tw loj (txog 40%) yam tsis muaj kev poob zoo. Tsis zoo li yav dhau los tsim tej yaam num siv tshuab kev kawm txoj kev rau suab compression, EnCodec tuaj yeem siv tsis tau tsuas yog rau kev hais lus ntim, tab sis kuj rau suab paj nruag compression nrog tus nqi piv txwv ntawm 48 kHz, sib xws rau qib ntawm audio CDs. Raws li cov neeg tsim tawm ntawm cov codec tshiab, thaum xa nrog ib qho bitrate ntawm 64 kbps piv rau MP3 hom, lawv muaj peev xwm nce qib ntawm lub suab compression los ntawm kwv yees li kaum npaug thaum tswj cov qib zoo tib yam (piv txwv li, thaum siv. MP3, yuav tsum muaj bandwidth ntawm 64 kbps, rau kev sib kis nrog qhov zoo tib yam hauv EnCodec yog txaus 6 kbps).

Lub codec architecture yog tsim nyob rau hauv ib tug neural network nrog ib tug "transformer" architecture thiab yog raws li plaub txuas: encoder, quantizer, decoder thiab discriminator. Lub encoder rho tawm cov tsis tau ntawm lub suab cov ntaub ntawv thiab hloov cov kwj packed mus rau tus ncej qis. Lub quantizer (RVQ, Residual Vector Quantizer) hloov cov kwj tawm los ntawm tus encoder rau hauv cov pob ntawv, compressing cov ntaub ntawv raws li qhov kev xaiv bitrate. Cov zis ntawm lub quantizer yog ib tug compressed sawv cev ntawm cov ntaub ntawv, haum rau kis tau tus mob nyob rau hauv lub network los yog txuag rau disk.

Tus decoder decodes lub compressed sawv cev ntawm cov ntaub ntawv thiab reconstructs tus thawj suab nthwv dej. Tus neeg ntxub ntxaug txhim kho qhov zoo ntawm cov qauv tsim, coj mus rau hauv tus account tus qauv ntawm tib neeg auditory xaav. Txawm hais tias qib zoo thiab bitrate, cov qauv siv rau kev encoding thiab decoding yog txawv los ntawm cov kev xav tau me me (xws li tsim nyog rau kev ua haujlwm ntawm lub sijhawm tiag tiag yog ua ntawm ib qho CPU core).

Facebook tshaj tawm EnCodec audio codec siv tshuab kev kawm


Tau qhov twg los: opennet.ru

Ntxiv ib saib