Audio AI JAGORA

HiFi-GAN dan GAN Vocoders

HiFi-GAN wani vocoder ne mai haɓaka-ƙiyayya wanda ke jujjuya mel-spectrogram zuwa ingantaccen yanayin sautin sauti kusan nan take, yana samar da ingantaccen magana mai inganci fiye da ainihin lokaci.

2 min karatuAn sabunta ta ƙarshe

Dubawa

Ya zama daidaitaccen mataki na ƙarshe na rubutu-zuwa-magana na zamani saboda yana da sauri, mara nauyi, da wuyar ganewa daga ainihin rikodin.

Zurfafa nutsewa

Vocoder shine mataki na ƙarshe a cikin mafi yawan bututun TTS: samfurin kamar Tacotron ko FastSpeech yana annabta mel-spectrogram (hoton ƙaramin hoto na mita akan lokaci), kuma vocoder ya cika ainihin samfuran waveform. Farkon vocoders na jijiyoyi kamar WaveNet sunyi sauti mai kyau amma sun haifar da samfurin sauti-by-samfuri, yana sa su jinkirin jinkiri. HiFi-GAN, wanda Kong, Kim, da Bae suka fitar a cikin 2020, sun maye gurbin wannan madauki na autoregressive tare da janareta na gaba da ciyarwa guda ɗaya wanda ya horar da gaba. Babban dabararsa ita ce ta amfani da masu wariya da yawa waɗanda ke yin hukunci da sauti a ma'auni daban-daban da kuma tsarin lokaci daban-daban, yana tilasta janareta don samun nau'ikan rubutu mai kyau da kuma yanayin lokaci daidai. Sakamakon shine 22 kHz magana da aka haɗa ɗaruruwan sau da sauri fiye da ainihin lokacin akan GPU, tare da ingantaccen sauti na gaskiya na ƙasa.

Fahimtar Fasaha

Generator na HiFi-GAN yana haɓaka misalan mel-spectrogram ta hanyar jujjuyawar juzu'i, tare da toshe shingen Filin Karɓar Maɗaukaki waɗanda ke haɗa nau'ikan kernel daban-daban da haɓaka don kama nau'ikan igiyoyin kalaman daban-daban. Iyalan masu nuna wariya guda biyu suna yin aikin ɗan sanda: Mai Wariya na Zamani da yawa yana sake fasalin siginar 1D zuwa grid na 2D a cikin manyan abubuwa kamar 2, 3, 5, 7, 11 don kama lokaci na lokaci, kuma Mai Bambancin Sikeli na Multi-Scale yana nazarin yanayin motsi a ƙayyadaddun ƙuduri da yawa. Mel-spectrogram da asarar madaidaicin fasali suna ci gaba da kwanciyar hankali.

Dabarun Tasiri

Shiga ku isa

Yana inganta samun dama ta hanyar rubutu, ba da labari, da mu'amalar murya.

Kudin da kasafin kuɗi

Ƙungiyoyin kafofin watsa labaru na iya jigilar sauti mai gogewa cikin sauri tare da ƙaramin kasafin kuɗi.

Gudu da sikelin

Tsarin fuskantar abokin ciniki na iya aiwatar da hulɗar magana a mafi girman ma'auni.

Makomar HiFi-GAN da GAN Vocoders

GAN vocoders suna ci gaba da ƙara ƙanƙanta da sauri: zuriya kamar BigVGAN suna ƙara ƙwaƙƙwaran ƙiyayya don haɗawa cikin mawaƙa, kayan kida, da harsunan da ba a gani ba, yayin da UnivNet da Vocos ke turawa zuwa ga duniya baki ɗaya. Bambance-bambancen yawo da kan na'urori yanzu suna gudanar da vocoding a cikin wayoyi da belun kunne don mataimakan rashin jinkiri. Ana ƙara haɓaka, watsawa da samfuran sauti masu dacewa da gudana zuwa cikin masu samar da hanyar wucewa ta hanyar GAN, suna haɗa amincin watsawa tare da saurin GAN. Yi tsammanin masu kada kuri'a za su ɓata cikin maƙasudi na gaba ɗaya-maƙasudin jigon sauti na jiwuwa masu ƙarfin magana da kiɗa.

Aiwatar da Gaskiyar Duniya

Samar da fitowar magana na mataimakan kama-da-wane da ƙa'idodin kewayawa waɗanda ke buƙatar martani ba tare da jinkiri ba.

Ƙaddamar da kullin murya na ainihin lokaci da kayan aikin dubbing inda aka sanya mel-spectrogram mai cloned zuwa sauti mai sauti na halitta.

Tuki littafin mai jiwuwa da dandamali na ba da labari na podcast waɗanda ke haɗa sa'o'i na magana cikin sauri da arha.

Yin hidima azaman matakin sigar igiyar ruwa a cikin masu haɗar murya da kiɗan kiɗa ta hanyar nau'ikan waƙoƙin duniya irin na BigVGAN.

Hatsari & Tsare-tsare

Rashin amfani da murya da haɗarin kwaikwaya yana ƙaruwa lokacin da aka rasa izini.

Daidaituwa na iya faɗuwa cikin lafuzza, yaruka, ko mahalli masu hayaniya.

Ana iya kuskuren sauti na roba don ingantacciyar magana ba tare da bayyananniyar lakabi ba.

Taswirar Hanya

1

Sami tabbataccen izini don ɗaukar murya, cloning, da sake amfani.

2

Gwajin ingantattun masu magana daban-daban da yanayin baya.

3

Ƙayyade lokacin da dole ne ɗan adam ya duba ko ya amince da abubuwan da aka fitar.

4

Yi lakabin sauti na roba da kuma adana bayanan da aka tabbatar don yin lissafi.

Ci gaba da Bincike

Free newsletter

Get the daily AI briefing

Three verified AI stories every weekday morning, written in plain English. Free forever, no ads.

One email each weekday. Unsubscribe in one click. We never sell or share your address.

Test yourself

Take the HiFi-GAN and GAN Vocoders quiz

Instant feedback on every answer, and a shareable certificate with a verifiable ID once you pass a course.

Fara tambayoyi

Support free AI education. AI Understanding is a 501(c)(3) nonprofit — no ads, no paywall, ever. Make a donation

Jagora na gaba

Daidaici WaveGAN Vocoder

Tambayoyin da ake yawan yi

Menene HiFi-GAN da GAN Vocoders?

HiFi-GAN wani vocoder ne mai haɓaka-ƙiyayya wanda ke jujjuya mel-spectrogram zuwa ingantaccen yanayin sautin sauti kusan nan take, yana samar da ingantaccen magana mai inganci fiye da ainihin lokaci. Ya zama daidaitaccen mataki na ƙarshe na rubutu-zuwa-magana na zamani domin yana da sauri, mara nauyi, da wuya a bambanta da ainihin rikodin.

Menene aikin farko na mai vocoder kamar HiFi-GAN a cikin bututun rubutu-zuwa-magana?

Vocoder shine mataki na ƙarshe wanda ke haɗa ainihin samfuran yanayin motsi daga mel-spectrogram wanda samfurin sauti ya samar.

Me yasa HiFi-GAN yayi sauri fiye da na farko WaveNet vocoder?

WaveNet ya haifar da samfurin sauti-by-samfuri (a kai tsaye), yayin da janareta na gaba-gaba na HiFi-GAN yana fitar da tsarin igiyar ruwa a cikin harbi ɗaya, yana samun saurin sauri fiye da ainihin-lokaci.

Menene wariya na Zamani da yawa na HiFi-GAN musamman yana taimakawa kama?

Mai Rarraba Tsakanin Zamani da yawa yana sake fasalin tsarin kalaman 1D zuwa 2D ta amfani da ƙididdiga masu ƙididdigewa don gano tsarin lokaci-lokaci da ke daure da farar.

GAN vocoders an horar da 'yan adawa.' Me hakan ke nufi a nan?

A cikin saitin GAN, janareta ya koyi samar da sifofin raƙuman ruwa na gaske don yaudarar hanyoyin sadarwar wariya waɗanda aka horar da su faɗi gaskiya daga sautin roba.

Wanne vocoder daga baya ya haɓaka HiFi-GAN don haɓaka mafi kyawun muryoyin da ba a gani, waƙa, da kayan kida?

BigVGAN yana haɓaka HiFi-GAN kuma yana ƙara kunnawa na lokaci-lokaci na anti-aliased, inganta haɓakawa ga sautin da ba a rarrabawa kamar kiɗa da kayan kida.