HAGAHA Audio AI

Qaab-dhismeedka Conformer

Conformer-ku waa xannibaadda neerfaha ee isku xidha is-fahamka iyo is-fiirsiga, oo soo qabanaya qaababka dhawaaqa maxaliga ah ee qaabaysan iyo macnaha fog ee hal lakab.

2 daqiiqo akhriMarkii u dambaysay ee la cusbooneysiiyay

Dulmar

It became the de facto standard encoder for state-of-the-art speech recognition.

quusid qoto dheer

Waxaa soo bandhigay Google sanadka 2020, Conformer-ku waxa uu ka jawaabay xiisad muhiim ah oo ku saabsan qaabaynta maqalka: is-fiirinta (ka Transformers) aad bay ugu fiican tahay macnaha guud laakiin way daciifsan tahay maxalli ahaan, qaabab quruxsan oo kala sooca telefoonnada, halka convolutions ay aad uga fiican yihiin gudaha laakiin halgan ugu jira in la arko hadal dheer. Qalabka Conformer-ku wuxuu ku tolaa naqshad 'sandwich' ah: qayb-tallaabo nus-tallaabo ah moduleka hore-u-qaadista, ka dib module is-fiirsi badan oo madax-madaxeed ah, ka dib moduleka convolution, ka dibna cutubka-hormarinta nuska-tallaabada labaad, oo leh lakabka caadiga ah iyo isku xirka haraaga oo dhan. Qaybta is-daba-marintu waxa ay isticmaashaa isku-dhafan oo si qoto dheer loo kala saari karo iyo unug toosan oo albaabbada loo xidhay. Iyaga oo is dhexgaya habka maxalli ah iyo kuwa caalamiga ah ee gudaha baloog kasta, Conformer encoders waxay jarayaan heerka khaladka kelmad si aad ah marka loo eego Transformer saafi ah ama sal-dhis sal-dhaxeed saafi ah oo ku salaysan halbeegyada sida LibriSpeech.

Aragtida Farsamada

Qaab dhismeedka 'Macaron' ee saxeexa wuxuu ku duubaa dareenka iyo is-waafajinta u dhexeeya laba lakab oo hore-u-quudineed ah, mid walbana wuxuu ku biirinayaa hadhaaga miisaan nuska ah (oo ah 0.5), oo ay dhiirigelisay falanqaynta lammaanaha Transformer FFN. Qaabka convolution sida caadiga ah waxa uu isku xidhaa isbedbedel toosan oo leh firfircooni GLU, isbadal qoto dheer, dufcaddii caadiga ahayd, firfircoonida Swish, iyo iswaydaarsiga micnaha ugu dambeeya - hab wax ku ool ah oo lagu qaabeeyo macnaha guud iyada oo aan qarxin tirinta cabbirka.

Saamaynta Istiraatijiyadeed

Helitaanka iyo gaarsiinta

Waxay wanaajisaa marin u helida iyada oo loo marayo qoraal-qorid, sheeko, iyo is-dhexgalyo cod.

Qiimaha iyo miisaaniyada

Kooxaha warbaahintu waxay ku soo rari karaan codka sifaysan si degdeg ah iyagoo wata miisaaniyado yaryar.

Xawaaraha iyo miisaanka

Nidaamyada u jeedda macmiisha waxay ka baaraandegi karaan isdhexgalka hadalka si weyn.

Mustaqbalka Conformer Architecture

Conformers hadda waxay u adeegaan sida codeeyaha laf-dhabarka ee beddelaha iyo CTC/fiirinta ASR, nashqadana waxay ku faaftay tarjumaada hadalka, aqoonsiga hadalka, iyo ogaanshaha dhacdada maqalka. Cilmi-baadhis fir fircoon ayaa sahlaysa dareenka maqalka dheer (fiiro toos ah iyo jeexjeexan oo loogu talagalay baahinta), waxay ka saartaa Conformers isticmaalka aaladda, waxayna ku lamaanaysaa tababbar iskiis u ilaalinayo. Kala duwanaanshiyaha sida Squeezeformer iyo Conformer Waxtarka leh ayaa sii riixaya saxsanaanta-ka-hortagga- xisaabinta ganacsiga-off.

Dhaqangelinta Adduunka-dhabta ah

U adeega sida codeeriyaha ee nidaamka ASR ee wax soo saarka ee ka dambeeya kaaliyeyaasha codka iyo kutubta

Awood siinta moodooyinka tarjumaadda hadalka ee qora oo tarjuma luqadda lagu hadlo dhammaad ilaa dhammaad

Laf-dhabar u ah xaqiijinta hadalka iyo shubanka, aqoonsiga qofka hadlay markii uu shir ku jiray

Dhacdada maqalka iyo kala soocida dhawaaqa, sida ogaanshaha alaarmiga, hadalka, ama muusiga qulqulka

Khatarta & Dariiqyada Ilaalada

Si xun u isticmaalka codka iyo khataraha is-yeelyeelku way kordhaan marka oggolaanshaha la waayo.

Saxnimadu waxay hoos ugu dhici kartaa lahjadaha, lahjadaha, ama jawiga buuqa badan.

Maqalka synthetic waxaa lagu khaldi karaa hadal dhab ah iyada oo aan si cad loo calaamadin.

Qorshe Hawleedka Dhaqangelinta

1

Hel ogolaansho cad oo ku saabsan qabashada codka, xidhitaanka, iyo dib u isticmaalka

2

Tijaabi tayada ku hadasha kala duwan iyo xaaladaha asalka.

3

Qeex marka bani'aadamku ay tahay inuu dib u eego ama oggolaado wax soo saarka.

4

Ku calaamadee codka synthetic oo xafid diiwaannada la-xisaabtanka.

Sii wad Sahaminta

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 Conformer Architecture quiz

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

Bilow kedis

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

Hagaha xiga

DeepSpeech Architecture

Su'aalaha soo noqnoqda

What is Conformer Architecture?

Conformer-ku waa xannibaadda neerfaha ee isku xidha is-fahamka iyo is-fiirsiga, oo soo qabanaya qaababka dhawaaqa maxaliga ah ee qaabaysan iyo macnaha fog ee hal lakab. Waxa ay noqotay de facto standard codeer ee aqoonsiga hadalka-casriga ah.

Waa maxay labada hab ee uu Conformer architecture ku daray hal block?

Conformer-ku waxa uu isku xidhaa isku-dhafan (qaababka maxaliga ah) oo leh is-fiirsasho (marka la eego macnaha guud) gudaha qayb kasta.

Waa maxay sababta isku-darka miyir-qabka iyo feejignaanta ay si gaar ah faa'iido ugu leedahay hadalka?

Isqabqabsigu waxa uu aad uga fiican yahay tilmaamaha maxaliga ah ee wanaagsan ee lagu kala soocayo taleefoonnada, halka is-fiirsigu uu ku tiirsan yahay dhammaan odhaahda; Conformer ayaa helaya labadaba.

Waa maxay qaabka 'Macaron' ama sanduuqa sanduuqa Conformer?

Conformer-ku waxa uu dhigayaa is-fiirsi iyo qaybaha iswaafajinta inta u dhaxaysa laba qaybood oo gudbin ah, mid walbana lagu dabaqay hadhaaga miisaan nuska ah.

Ururkee ayaa soo bandhigay Conformer, iyo sanadkee?

Conformer-ka waxa keenay Google cilmi-baarayaal 2020 oo si degdeg ah u noqday cod-garashada hadalka caadiga ah.

Maxaa ka mid ah cutubka isdhaafsiga ee ku dhex jira Conformer sida caadiga ah?

Module convolution wuxuu ku xidhaa iskudhafan dhinac dhinac u jeeda oo leh unug toosan oo xidhan, iswaydaarsi qoto dheer, caadiyeyn dufcaddii, iyo hawl qabad Swish, oo ku dhamaanaya wareeg kale oo dhinac u jeeda.