EnCodec, iyo itsva Meta odhiyo codec

encodec

Encodec icodec inodhinda ichishandisa neural network ine compression rate inokwana gumi

Munguva pfupi yapfuura, Meta (yaimbova Facebook) yakafumura codec yayo nyowani inonzi EnCodec, , que inoshandisa michina yekudzidza michina kuwedzera compression reshiyo pasina kurasikirwa kunaka.

Iyo nzira nyowani inogona kudzvanya uye decompress odhiyo munguva chaiyo kuti uwane mamiriro-e-iyo-anyanzvi saizi kuderedza. iyo codec inogona kushandiswa kune ese ari maviri ekutepfenyura odhiyo munguva chaiyo seye encoding yekuchengetedza gare gare mumafaira.

Nhasi, tiri kudonongodza mafambiro aita Yako AI Research (FAIR) munzvimbo yeAI-powered audio hyper-compression. Fungidzira uchiterera meseji yekuteerera yeshamwari munzvimbo isina kubatana uye isina kumira kana kubondera. Tsvagiridzo yedu inoratidza mashandisiro atingaita AI kutibatsira kuita izvi.

InCodec kupa mhando mbiri ready to download:

  1. A causal modhi inoshandisa 24 kHz sampuli mwero, inotsigira monophonic audio chete, uye inodzidziswa pane dzakasiyana siyana data redhiyo (yakakodzera encoding yekutaura). Iyo modhi inogona kushandiswa kurongedza odhiyo data yekufambisa pabit rates ye1,5, 3, 6, 12 uye 24 kbps.
  2. Iyo isiri-causal modhi inoshandisa 48kHz sampuli mwero, inotsigira stereo ruzha, uye yakadzidziswa mimhanzi chete. Iyo modhi inotsigira bit mitengo ye3, 6, 12 uye 24 kbps.

Pamuenzaniso wega wega, modhi yekuwedzera yemutauro yakagadzirirwa, chii inobvumira kuwedzera kukuru muyero yekumanikidza (kusvika ku40%) pasina kurasikirwa kwemhando. Kusiyana nemapurojekiti apfuura ekushandisa matekinoroji ekudzidza muchina kuodhiyo yekumanikidza, EnCodec inogona kushandiswa kwete kungorongedza kutaura, asiwo yekumanikidza mimhanzi ine sampling frequency ye48 kHz, inoenderana nehuwandu hwemaCD ekuteerera.

Sekureva kwevagadziri veiyo codec nyowani, nekufambisa nechiyero chidiki che64 kbps zvichienzaniswa neiyo MP3 fomati, vakakwanisa kuwedzera redhiyo yekumisikidza reshiyo kanenge kagumi vachichengeta chiyero chakafanana chemhando (semuenzaniso, pakushandisa MP3. zvinoda bandwidth ye64 kbps, kuendesa nemhando imwechete muEnCodec, 6 kbps yakakwana).

Iyi data inogona kuzogadziriswa uchishandisa neural network. Tawana chiyero che10x chekudzvanya zvichienzaniswa neMP3 pa64kbps, pasina kurasikirwa kwemhando. Nepo hunyanzvi uhu hwakamboongororwa pakutaura, ndisu vekutanga kuita kuti ishandire 48 kHz sampled stereo audio (kureva CD mhando), inova chiyero chekugovera mimhanzi.

Chimiro chekodheki Iyo yakavakirwa pahwaro hweiyo neural network ine "shanduko" yekuvaka uye inobva pazvisungo zvina: encoder, quantizer, decoder uye kusarura:

  • El encoder inobvisa ma paramita kubva ku data yezwi uye inoishandura kuita rwizi rwakazara pakiti yakaderera furemu.
  • El quantifier (RVQ, Residual Vector Quantizer) inoshandura iyo encoder inobuda rwizi kuita seti yemapaketi, ichimanikidza ruzivo rwune hukama kune yakasarudzwa bit rate. Iyo yakabuda ye quantizer inomiririra yakamanikidzwa yedata yakakodzera kutapurirana pamusoro petiweki kana kuchengetedza kune dhisiki.
  • El decoder inosarudza iyo yakamanikidzwa data inomiririra uye inovakazve yekutanga ruzha wave.
  • El rusarura inovandudza kunaka kwemasampuli akagadzirwa (sample) tichifunga nezvemuenzaniso wekunzwa kwemunhu.

Pasinei nechiyero chemhando uye bitrate, mamodheru anoshandiswa encoding uye decoding anosiyana mune zvine mwero zvinodiwa zviwanikwa (maverengero anodiwa pakuita-chaiyo-nguva oparesheni anoitwa pane imwechete CPU musimboti).

Chekupedzisira, kune avo vamuri kufarira, munofanirwa kuziva kuti iyo referensi yekuitwa kweEnCodec yakanyorwa muPython uchishandisa iyo PyTorch chimiro uye inopihwa rezinesi pasi peCC BY-NC 4.0 (Creative Commons Attribution-NonCommerce) rezinesi rekushandisa zvisiri zvekutengesa. chete.

Kana iwe uchida kudzidza zvakawanda nezvazvo, unogona kubvunza ruzivo pa chinotevera chinongedzo.


Siya yako yekutaura

Your kero e havazobvumirwi ichibudiswa. Raida minda anozivikanwa ne *

*

*

  1. Inotarisira data: AB Internet Networks 2008 SL
  2. Chinangwa cheiyo data: Kudzora SPAM, manejimendi manejimendi.
  3. Legitimation: Kubvuma kwako
  4. Kutaurirana kwedata
  5. Dhata yekuchengetedza: Dhatabhesi inobatwa neOccentus Networks (EU)
  6. Kodzero: Panguva ipi neipi iwe unogona kudzora, kupora uye kudzima ruzivo rwako