UMHLAHLANDLELA WE-AI womsindo

I-Conformer Architecture

I-Conformer iyibhulokhi yenethiwekhi ye-neural ehlanganisa i-convolution nokuzinaka, ithwebula womabili amaphethini omsindo wendawo anohlamvu oluhle kanye nomxholo webanga elide kusendlalelo esisodwa.

Uhlolojikelele

I-Conformer iyibhulokhi yenethiwekhi ye-neural ehlanganisa i-convolution nokuzinaka, ithwebula womabili amaphethini omsindo wendawo anohlamvu oluhle kanye nomxholo webanga elide kusendlalelo esisodwa. Ibe isishumeki esijwayelekile se-de facto sokubonwa kwenkulumo yesimanje.

I-Conformer Architecture ihlezi ku-audio-AI workflows eguqula inkulumo, umculo, nomsindo wokuxhumana, ukufinyeleleka, nokukhiqizwa kwemidiya.

I-Deep Dive

Yethulwe ngu-Google ngo-2020, i-Conformer yaphendula ukungezwani okubalulekile ekwenzeni imodeli yomsindo: ukuzinaka (kusuka ku-Transformers) kuhle kakhulu emhlabeni jikelele kodwa kubuthakathaka emaphethini endawo, ahlelwe kahle ahlukanisa amafoni, kuyilapho ukuguquguquka kugqama endaweni kodwa kunzima ukubona iphimbo elide. Ibhulokhi ye-Conformer iwahlanganisa ndawonye ngomklamo 'wesemishi': imojuli yesinyathelo esiyisa phambili sokuphakelayo, bese kuba imojuli yokunaka enamakhanda amaningi, bese kuba imojula ye-convolution, bese kuba imojuli yesinyathelo sesibili sokuphakelayo, enokwejwayelekile kwesendlalelo noxhumo oluyinsalela kulo lonke. Imojula ye-convolution isebenzisa ama-convolutions ahlukanisekayo ajulile kanye neyunithi yomugqa enesango. Ngokuxhumanisa ukucubungula kwendawo nembulunga yonke ngaphakathi kwebhulokhi ngayinye, izifaki khodi ze-Conformer zisika izilinganiso zephutha lamagama kakhulu nge-Transformer emsulwa noma imigqa eyisisekelo ye-convolutional kumabhentshimakhi afana ne-LibriSpeech.

I-Technical Insight

Isakhiwo sesiginesha 'I-Macaron' sisonga ukunaka nokuguqulwa phakathi kwezendlalelo ezimbili zokuphakela phambili, ngayinye inikela ngensalela enesisindo esiyingxenye (isici esingu-0.5), esigqugquzelwe ukuhlaziya kwamapheya e-Transformer FFN. Imojula ye-convolution ivamise ukuhlanganisa i-convolution eqondile ngokuvula i-GLU, ukuguquguquka okujulile, ukwenziwa kwe-batch normalization, ukwenziwa kusebenze kwe-Swish, kanye ne-convolution yokugcina ephusile - indlela ephumelelayo yokumodela umongo wendawo ngaphandle kokuqhuma ukubala kwepharamitha.

I-Mastering Conformer Architecture

I-Conformer iyibhulokhi yenethiwekhi ye-neural ehlanganisa i-convolution nokuzinaka, ithwebula womabili amaphethini omsindo wendawo anohlamvu oluhle kanye nomxholo webanga elide kusendlalelo esisodwa. Ibe isishumeki esijwayelekile se-de facto sokubonwa kwenkulumo yesimanje. I-Conformer Architecture ihlezi ku-audio-AI workflows eguqula inkulumo, umculo, nomsindo wokuxhumana, ukufinyeleleka, nokukhiqizwa kwemidiya. Ukuze wakhe ukuqonda okujulile, phatha i-Conformer Architecture njengemodeli yokusebenza, hhayi isici esisodwa: chaza imiphumela efiselekayo, ucacise ukucabanga, futhi uhlukanise lokho uhlelo olungakwenza ngokwethembeka kulokho okusadinga ukwahlulela kochwepheshe.

Empeleni, amaqembu aqinile asebenzisa i-Conformer Architecture aphatha ikhwalithi, ukubambezeleka, kanye nemvume njengezingxenye ezibalulekile ngokulinganayo zesu lokuthunyelwa. Babhala imibandela yempumelelo ecacile, ukuhlola okuqhathaniswa nedatha engokoqobo nokugeleza komsebenzi, futhi baphindaphinde ngokusekelwe kumaphethini okuhluleka aqashiwe esikhundleni sokuwina kwebhentshimakhi yesikhathi esisodwa. Yilapho ukuqonda kwethiyori kuguquka kube amandla ahlala njalo kuwo wonke umkhiqizo, inqubomgomo, kanye nokusebenza.

Ithuthukisa ukufinyeleleka ngokuloba, ukulandisa, nezixhumi ezibonakalayo zezwi. Ngesikhathi esifanayo, ukusetshenziswa kabi kwezwi kanye nezingozi zokuzenza ongeyena ziyakhuphuka uma imvume ingekho. Indlela eqine kakhulu iwukuhlanganisa isivinini sokuhlola nesiyalo sokuphatha: qhuba abashayeli bezindiza, bamba ubufakazi, ushicilele amalogi ezinqumo, futhi ubuyekeze izivikelo ngokuqhubekayo njengoba imodeli yokuziphatha, okulindelwe ngabasebenzisi, kanye nezimfuneko zokulawula zishintsha.

I-Strategic Impact

Ithuthukisa ukufinyeleleka ngokuloba, ukulandisa, nezixhumi ezibonakalayo zezwi.

Ithuthukisa ukufinyeleleka ngokuloba, ukulandisa, nezixhumi ezibonakalayo zezwi. Ekusetshenzisweni kwekhwalithi ephezulu, lokhu kuhunyushwa emithethweni yokusebenza elinganisekayo, imingcele yobunikazi, nemikhuba yokubuyekeza ephindelelayo ukuze amaqembu akwazi ukukala ukuzethemba esikhundleni sokukala ukungaqondakali.

Amaqembu emidiya angathumela umsindo opholishiwe ngokushesha ngamabhajethi amancane.

Amaqembu emidiya angathumela umsindo opholishiwe ngokushesha ngamabhajethi amancane. Ekusetshenzisweni kwekhwalithi ephezulu, lokhu kuhunyushwa emithethweni yokusebenza elinganisekayo, imingcele yobunikazi, nemikhuba yokubuyekeza ephindelelayo ukuze amaqembu akwazi ukukala ukuzethemba esikhundleni sokukala ukungaqondakali.

Amasistimu abhekene nekhasimende angacubungula ukusebenzelana okukhulunyiwe ngesilinganiso esikhulu.

Amasistimu abhekene nekhasimende angacubungula ukusebenzelana okukhulunyiwe ngesilinganiso esikhulu. Ekusetshenzisweni kwekhwalithi ephezulu, lokhu kuhunyushwa emithethweni yokusebenza elinganisekayo, imingcele yobunikazi, nemikhuba yokubuyekeza ephindelelayo ukuze amaqembu akwazi ukukala ukuzethemba esikhundleni sokukala ukungaqondakali.

Ikusasa le-Conformer Architecture

Izikhombi manje zisebenza njengesishumeki esiwumgogodla se-transducer kanye ne-CTC/attention ASR, futhi idizayini isisabalele ekuhumusheni inkulumo, ukubonwa kwesipikha, nokutholwa komcimbi womsindo. Ucwaningo olusebenzayo luqondisa ukunaka komsindo omude (ukunaka okuqondile nokucutshungulwayo ukuze kusakazeke), kudidiyela ama-Conformers ukuze asetshenziswe kudivayisi, futhi iwabhangqa nokuziqeqesha kwangaphambili okuzigadile. Okuhlukile okufana ne-Squeezeformer kanye ne-Efficient Conformer kuqhuba ukuhwebelana kokunemba kuqhathaniswa nekhompiyutha ngokuqhubekayo.

Ukuqaliswa Komhlaba Wangempela

Isebenza njengesifaki khodi ekukhiqizeni ukusakaza izinhlelo ze-ASR ngemuva kwezisizi zezwi nokubizela

Inika amandla amamodeli okuhumusha inkulumo abhala futhi ahumushe ulimi olukhulunywayo ekugcineni ukuya ekugcineni

Umgogodla wokuqinisekisa isipikha kanye nokufakwa kwedayari, ukuhlonza ukuthi ubani okhulume lapho usemhlanganweni

Umcimbi womsindo nokuhlukaniswa komsindo, njengokuthola ama-alamu, inkulumo, noma umculo okusakazweni

Amaphethini Okusebenzisa

Conformer Architecture in practice

Isebenza njengesifaki khodi ekukhiqizeni ukusakaza izinhlelo ze-ASR ngemuva kwezisizi zezwi nokubizela.

Isebenza njengesifaki khodi ezinhlelweni zokusakaza ezisakazwayo ze-ASR ngemuva kwabasizi bezwi namaQembu okubizela ngokuvamile athola imiphumela engcono uma echaza imingcele yekhwalithi ngaphambili, egcina indlela yokukhuphuka yabantu yamakesi asemaphethelweni, futhi alandelele kokubili izinzuzo zokukhiqiza nezindleko zamaphutha ngokuhamba kwesikhathi.

Conformer Architecture in practice

Inika amandla amamodeli okuhumusha inkulumo abhala futhi ahumushe ulimi olukhulunywayo ekugcineni ukuya ekugcineni.

Amandla amamodeli okuhumusha inkulumo aloba futhi ahumushe ulimi olukhulunywayo kusukela ekupheleni kuye ekupheleni Amaqembu ngokuvamile athola imiphumela engcono uma echaza izilinganiso zekhwalithi ngaphambili, egcina indlela yokukhuphuka yabantu yamakesi asemaphethelweni, futhi alandelela kokubili izinzuzo zokukhiqiza nezindleko zamaphutha ngokuhamba kwesikhathi.

Conformer Architecture in practice

Umgogodla wokuqinisekisa isipikha kanye nokufakwa kwedayari, ukuhlonza ukuthi ubani okhulume lapho usemhlanganweni.

Umgogodla wokuqinisekisa isipikha nokufakwa kwedayari, ukuhlonza ukuthi ubani owakhuluma lapho esemhlanganweni Amathimba ngokuvamile athola imiphumela engcono lapho echaza izilinganiso zekhwalithi ngaphambili, agcina indlela yokukhuphuka komuntu yamakesi asemaphethelweni, futhi alandelele kokubili izinzuzo zokukhiqiza nezindleko zamaphutha ngokuhamba kwesikhathi.

Conformer Architecture in practice

Umcimbi womsindo nokuhlukaniswa komsindo, njengokuthola ama-alamu, inkulumo, noma umculo okusakazweni.

Umcimbi womsindo nokuhlukaniswa komsindo, okufana nokuthola ama-alamu, inkulumo, noma umculo ekusakazeni Amathimba ngokuvamile athola imiphumela engcono uma echaza ikhwalithi ephezulu ngaphambili, egcina indlela yokukhuphuka yabantu yamakesi asemaphethelweni, futhi alandelele kokubili izinzuzo zokukhiqiza nezindleko zamaphutha ngokuhamba kwesikhathi.

Izingozi & Guardrails

!

Ukusetshenziswa kabi kwezwi kanye nezingozi zokuzenza ongeyena ziyanda uma imvume ingekho.

!

Ukunemba kungase kwehle kuzo zonke izinhlobo zokuphimisela, izilimi zesigodi, noma izindawo ezinomsindo.

!

Umsindo wokwenziwa ungenziwa iphutha njengenkulumo eyiqiniso ngaphandle kokulebula okucacile.

Ukuqalisa Umhlahlandlela

1

Thola imvume esobala yokuthwebula izwi, ukuhlanganisa, nokusebenzisa kabusha.

Thola imvume esobala yokuthwebula izwi, ukuhlanganisa, nokusebenzisa kabusha. Phatha isinyathelo ngasinye njengesango lobufakazi: uma imibandela ingafinyelelwa, misa ukukhishwa, vala igebe, bese unweba ukusetshenziswa.

2

Ikhwalithi yokuhlola kuzo zonke izipikha nezimo zangemuva.

Ikhwalithi yokuhlola kuzo zonke izipikha nezimo zangemuva. Phatha isinyathelo ngasinye njengesango lobufakazi: uma imibandela ingafinyelelwa, misa ukukhishwa, vala igebe, bese unweba ukusetshenziswa.

3

Chaza ukuthi kunini lapho umuntu kufanele abuyekeze noma agunyaze okuphumayo.

Chaza ukuthi kunini lapho umuntu kufanele abuyekeze noma agunyaze okuphumayo. Phatha isinyathelo ngasinye njengesango lobufakazi: uma imibandela ingafinyelelwa, misa ukukhishwa, vala igebe, bese unweba ukusetshenziswa.

4

Lebula umsindo wokwenziwa futhi ugcine amarekhodi atholakalayo ukuze aziphendulele.

Lebula umsindo wokwenziwa futhi ugcine amarekhodi atholakalayo ukuze aziphendulele. Phatha isinyathelo ngasinye njengesango lobufakazi: uma imibandela ingafinyelelwa, misa ukukhishwa, vala igebe, bese unweba ukusetshenziswa.

Qhubeka Uhlole