Ukuqaphela i-accent: Ukuvala igebe lolimi

• About the Author:
• Igama lomlobi

  quantumrun Foresight
• February 19, 2024

Isifinyezo sokuqonda

Ucwaningo lokuqaphela i-accent seluzuze ukubaluleka muva nje njengoba luzama ukuthuthukisa ukuxhumana kuzo zonke izilimi. Ubuchwepheshe bokuqaphela iphimbo lenkulumo (SAR) sekulungele ukuthuthukisa ukuxhumana kwamasiko ahlukene, ukunikeza ulwazi lomuntu siqu lokufunda, nokudala amathuba emisebenzi kuyilapho kuphakamisa imibuzo mayelana nobumfihlo bedatha kanye nokusetshenziswa kokuziphatha. Ukuthuthukiswa kwe-SAR kunemiphumela efinyelela kude, kusukela ekusizeni ukubambisana komhlaba wonke ekukhuthazeni ukufakwa komphakathi kanye nokuthuthukisa izinsizakalo eziphuthumayo.

Umongo wokubona i-accent

Ucwaningo lokuqaphela i-accent, olubaluleke kakhulu eminyakeni yamuva nje, lubandakanya izifundo ezibanzi kuzo zonke izilimi ezihlukahlukene ukuze kuthuthukiswe ukusebenza kwesistimu. Njengoba izinkampani eziningi zifaka imali ekunikeni amandla ukuhumusha kwesikhathi sangempela kuzo zonke izindlela ezihlukene, le ndawo yocwaningo ibilokhu izuza amandla. Isibonelo, ucwaningo lwango-2022 olushicilelwe ku-Arabian Journal for Science and Engineering lwasebenzisa i-convolutional neural networks (CNN), imodeli yokufunda ejulile (DL), kusetshenziswa izithombe ze-spectrogram ukuze kube lula ukukhishwa kwesici kumasignali omsindo (izingxoxo zesiNgisi saseBrithani). Ukunemba kwesistimu yokuqaphela iphimbo bekuphawuleka, ngokunemba okungamaphesenti angama-92.92 okuhlolwa okuncike ebulilini kanye namaphesenti angama-93.38 okuhlolwa okuncike ebulilini. 

Olunye ucwaningo lwango-2022 olushicilelwe ku-SSRN lubhekane nesidingo sokunemba okuphezulu kokubhala ezinhlelweni zokuqashelwa kwenkulumo okuzenzakalelayo (ASR), ikakhulukazi ezikhulumi ezingezona ezomdabu nezinezwi. Ucwaningo lugxile ekuboneni amaphimbo nokucebisa idathasethi yokuqeqeshwa ngedatha yenkulumo ene-accented eyahlukene ukuze kuthuthukiswe ukusebenza kwe-ASR. Kubandakanya i-prosodic (isigqi, umculo, nephimbo lenkulumo), izici zenkulumo yephimbo, nokushumeka kwesipikha kwathuthukisa ukunemba kwayo yonke imodeli futhi kwasizwa ekunanzeni kwephimbo okungelona elomdabu, kusetshenziswa idathasethi yangokwezifiso emboza izikhulumi zomhlaba wonke ezinamagama ahlukahlukene.

Ekugcineni, ucwaningo lwango-2024 lwalugxile ekuthuthukiseni i-Speech Accent Recognition (SAR) kusetshenziswa ukudlulisa ukufunda okuvela emisebenzini eyahlukene yokucubungula inkulumo. Ucwaningo lubonise ukuthi ukudlulisa ulwazi kusuka kumamodeli we-ASR kukhulisa kakhulu ukunemba kwe-SAR, ngentuthuko ehlobene ngamaphesenti angama-46.7. Ucwaningo lusebenzise i-Architecture ye-Conformer (imodeli ye-DL esetshenziswa ekucubungulweni kwenkulumo nomsindo) kanye nokuhlola kudathasethi ye-Vietnamese, okuveza ukusebenza kahle kwale ndlela. Sekukonke, lolu cwaningo lugqamise amandla okufunda okudluliswayo ukuze kuthuthukiswe ukuqashelwa kwephimbo ngezilimi ezinezinsizakusebenza eziphansi.

Umthelela ophazamisayo

Imizamo yokuthuthukisa ubuchwepheshe be-SAR isho ukuxhumana okubandakanyayo kanye nokusebenza kahle nobuchwepheshe. Abantu abavela ezizindeni ezihlukene zezilimi bangathola ukunemba nokuqonda okuthuthukisiwe lapho bexhumana namasistimu alawulwa ngezwi. Lokhu kuthambekela kungase kuthuthukise ukufinyeleleka, kuqinisekise ukuthi ubuchwepheshe buvumelane kakhulu nabantu abanezimpawu zokuphimisa ezihlukahlukene nezindlela zokukhuluma, ekugcineni kuvalwe izikhala zokuxhumana.

Izinkampani zingase zidinge ukubeka phambili ukuhlanganisa ubuchwepheshe bokuqaphela iphimbo lenkulumo kusevisi yabo yamakhasimende namasu okumaketha. Ngokwenza kanjalo, bangahlinzeka ngokusebenzisana kwamakhasimende okwenzelwe umuntu siqu nakakhulu, okubenza bakwazi ukubhekana kangcono nezidingo zendawo. Ukwengeza, amabhizinisi angase asebenzise lobu buchwepheshe ukuze athole imininingwane ejulile mayelana nezintandokazi nokuziphatha kwamakhasimende, okuvumela ukwenziwa kwesinqumo okwengeziwe okutholwe yidatha kanye nokunikezwa kwemikhiqizo okuthuthukisiwe.

Ohulumeni, nabo, bangazuza ekuthuthukisweni kobuchwepheshe be-SAR. Izinsiza zomphakathi zingasebenza ngempumelelo kakhulu ekukhonzeni imiphakathi ekhuluma izilimi eziningi, ukuqinisekisa ukuthi izakhamuzi ezivela ezizindeni ezihlukene zingakwazi ukuthola ulwazi olubalulekile lukahulumeni kanye nezinsizakalo. Ngaphezu kwalokho, lobu buchwepheshe bungaba nezicelo zokuphepha nezomthetho zokuhlaziya izwi nokuhlonza, okungase kuthuthukise imizamo yokuphepha yomphakathi.

Imithelela yokubonwa kwephimbo

Imithelela ebanzi yokuqashelwa kwephimbo ingase ihlanganise: 

• Ukuxhumana okushelelayo kwamasiko ahlukahlukene, okuzuzisa amabhizinisi omhlaba kanye nokukhuthaza ukusebenzisana komhlaba wonke.
• Okuhlangenwe nakho kokufunda okubandakanyayo nokomuntu siqu kwabafundi abanezimpawu ezihlukile nezizinda zolimi, okunciphisa ukungafani kwezemfundo.
• Izinkampani zijwayelanisa amasu azo okumaketha ukuze zifake ukukhangisa okwazi ukuphimisa, okubavumela ukuthi baxhumane nabathengi ngezinga lomuntu siqu futhi baqondise izibalo zezilimi ezithile.
• Imithetho yokuvikela ubumfihlo bedatha yezwi, ukubhekana nokukhathazeka okungaba khona mayelana nokuvikeleka kwedatha kanye nokusetshenziswa kokuziphatha kubuchwepheshe be-SAR.
• Amathuba emisebenzi kubuchwepheshe bolimi, isichasiselo sedatha, nokuthuthukiswa kwamamodeli.
• Amasevisi ezimo eziphuthumayo athuthukisiwe ngokuhlonza ngokunembile ulimi nokuphimisa kwabafonayo abacindezelekile, ukunika amandla izimpendulo ezisheshayo nezisebenza ngempumelelo.
• Abasizi bezwi bahlomeke ukunakwa okuphimisayo ukuze bathuthukise ukusebenzelana kwezakhamizi, ukufinyelela kumasevisi omphakathi, kanye nokuxhumana nomphakathi.
• Ukufakwa komphakathi kunciphisa ukucwasa ngokwezilimi kanye nokuchema ezimweni ezihlukahlukene zomphakathi.

Imibuzo okufanele icatshangelwe

• Ubuchwepheshe be-SAR bungakusiza kanjani emsebenzini wakho?
• Yiziphi izimiso zokuziphatha okufanele amabhizinisi nohulumeni bazicabangele lapho besebenzisa idatha ehlobene nokuphimisa ukuze kuthathwe izinqumo nokusebenzisa inqubomgomo?