I-Hy ASR 3.0 preview – imodeli yesizukulwane esisha yokuqaphela inkulumo eyethulwe yi-Tencent Hunyuan Uyasebenza

I-Hy ASR 3.0 preview iyimodeli yesizukulwane esisha yokuqaphela inkulumo eyethulwe yi-Tencent Hunyuan, eyakhelwe phezu kwemodeli yolimi enkulu i-Hy3, ihlanganisa ukuqaphela inkulumo okunembe kakhulu nokuqonda okujulile kwencazelo. Le modeli isekela isiShayina, isiNgisi, isiCantonese kanye nezifunda ezingu-10 zezilimi zesigodi, futhi inamandla okulungisa amaphutha ngokuzenzakalelayo ngokusekelwe kumongo, ukufaka amagama abalulekile nokusebenza ezimweni eziyinkimbinkimbi ezinjengomsindo omkhulu nokuhleba. Kuqoqo lokuhlola elivulekile, i-WER igcinwa cishe ku-3%, kanti iqoqo lokuhlola elizakhelwe ngaphakathi lihola izimbangi kuzo zonke izici.

Ukuqaphela okujwayelekile nokunembile: Isekela isiMandarin esijwayelekile, isiNgisi, isiCantonese nokuxubana kwesiShayina nesiNgisi, inciphise izinhlamvu ezingalungile, izinhlamvu ezingekho kanye nokwanda kwamaphutha kumarekhodi omsindo amade.
Ukuqonda incazelo yomongo: Isebenzisa i-Context ukuze ibambe ngokunembile umongo womsebenzisi, ilungise ngokuzenzakalelayo amagama aphinyiswa ngokufanayo futhi isuse ukudideka kwencazelo.
Ukuzivumelanisa nokufakwa kwamagama abalulekile: Isekela ukuqiniswa kwamagama abalulekile afana namagama emikhiqizo, amagama abantu namagama emboni, yehlise izindleko zokuxhumanisa nokugcina izinhlelo zebhizinisi.
Ukuqina ezindaweni eziyinkimbinkimbi: Yenzelwe ngokukhethekile izimo zomsindo ezinjengomsindo omkhulu, ukuhleba nokukhuluma kancane, igcine ukusebenza okuqinile kokuqaphela.
Ukusabalala okubanzi kwezilimi zesigodi: Isekela izifunda ezingu-10 zezilimi zesigodi kanye nezindawo ezingaphansi ezidlula ku-20, okuhlanganisa isiMandarin saseNyakatho-mpumalanga, isiCantonese, isiWu nesiMinnan.

Ukuthuthukiswa kwesakhiwo: Kusetshenziswa isakhiwo se-MoE, futhi imodeli eyisisekelo ithuthukiselwe ku-Hy3; kwakhiwe i-Encoder yezwi engadingi ukugadwa, eqeqeshwe ngedatha yezwi yamahora angamashumi ezigidi ukuze ikhiphe izethulo ezisezingeni eliphezulu zezimpawu zomsindo.
I-Scaling yokuqeqeshwa kwangaphambili: I-Encoder ne-LLM kuqeqeshwa ngokuhlanganyela, kufakwe idatha yezwi evela emithonjeni eminingi yamahora angamashumi ezigidi, okuhlanganisa izincwadi ezilalelwayo, izingxoxo, amavidiyo nomsakazo. Ngemijikelezo eminingi yokufakwa kwamakhono, kutholakala ukuqonda izilimi zesigodi namandla okumodela umongo.
Ukuqiniswa kwangemva kokuqeqeshwa: Kwakhiwe i-SFT Recipe esezingeni eliphezulu ehlanganisa umsindo oyinkimbinkimbi, umongo we-Any Context namakhono ajwayelekile; ukufunda kokuqinisa kwezigaba eziningi kusetshenziselwa ukuthuthukisa ngokwehlukana ukunemba kokuguqula inkulumo okujwayelekile, ukuncika komongo webanga elide nokusebenza ezimweni eziyinkimbinkimbi ezingavamile.

Xhuma nge-Tencent Cloud API noma uvule i-Tencent Yuanbao, bese ucindezela ubambe inkinobho yokukhuluma ukuze uyihlole mahhala.

Ukuthuthukiswa kokuqonda incazelo: Ngokusekelwe kumodeli yolimi enkulu i-Hy3, ihlanganisa ukuqaphela inkulumo okunembe kakhulu nokuqonda okujulile kwencazelo, iguquke ekuguquleni amagama ngamanye iye ekuqondeni umongo nasekukhipheni umphumela oqondile ngokuchofoza okukodwa.
Ukuhola okuphelele ekuhlolweni: Kuqoqo lokuhlola elivulekile, i-WER igcinwa cishe ku-3%; emaqoqweni okuhlola azakhelwe ngaphakathi, idlula izimbangi ezimweni ezijwayelekile, zezilimi zesigodi, zomongo nezomsindo eziyinkimbinkimbi.
Ukulungiswa kwamaphutha okuhlakaniphile ngokomongo: Ihlanganisa i-Context ukuze ibambe umongo ngokunembile, ilungise amagama aphinyiswa ngokufanayo, isuse ukudideka kwencazelo futhi isekele ukumodela kokuncika kwebanga elide.
Ukuzivumelanisa okusheshayo namagama abalulekile: Isekela ukufakwa kwamagama abalulekile afana namagama emikhiqizo, amagama abantu namagama emboni, yehlise izindleko zokuxhumanisa nokugcina izinhlelo zebhizinisi ezikhethekile.
Ukusabalala kwezilimi zesigodi nezindawo eziyinkimbinkimbi: Isekela izifunda ezingu-10 zezilimi zesigodi kanye nezindawo ezingaphansi ezidlula ku-20, igcine ukusebenza okuqinile ezindaweni eziyinkimbinkimbi zomsindo ezinjengomsindo omkhulu nokuhleba.

Isevisi yamakhasimende ehlakaniphile: Ezimweni zesevisi yamakhasimende ngocingo naku-inthanethi, i-Hy ASR 3.0 preview ibona ngokunembile amagama emikhiqizo namagama ebhizinisi ngokufakwa kwamagama abalulekile, yenze ukuguqulwa kwenkulumo ngesikhathi sangempela futhi ithuthukise kakhulu ukusebenza kahle kokuphendula kwesevisi.
Ukudala okuqukethwe: Ekukhiqizeni okuqukethwe komsindo omude njengama-podcast, imihlangano nezingxoxo, le modeli isebenzisa ukuqonda incazelo yomongo ukuze isuse ngokuzenzakalelayo ukudideka, ikhiphe umbhalo ohambisanayo ofundekayo nosezingeni eliphezulu ngokuchofoza okukodwa futhi yehlise izindleko zokuhlela kamuva.
Ukusesha ngezwi: Isekela abasebenzisi abasebenzisa izwi lezilimi zesigodi ekusesheni, futhi ibone ngokunembile imiyalo yezwi ezindaweni ezinomsindo, yehlise umgoqo wokufaka kubasebenzisi bezilimi eziningi nabanezindlela ezihlukahlukene zokuphimisa.
Ukusebenzisana ehhovisi: Ezimweni zokuguqula amarekhodi emihlangano ngesikhathi sangempela, le modeli igcina ukuqaphela okuqinile ngisho nalapho ibhekene nokufakwa kwezwi eliphansi njengokuhleba nokukhuluma kancane, isize amaqembu ekuxhumaneni kahle nasekugcineni ulwazi.
Amadivayisi ahlakaniphile: Ekusebenzisaneni namadivayisi e-IoT afana nokuzulazula ezimotweni namakhaya ahlakaniphile, le modeli yenzelwe ngokukhethekile izindawo ezinomsindo omkhulu, iqinisekise ukuqina kokuqaphela inkulumo nolwazi oluhle lomsebenzisi ezimweni eziyinkimbinkimbi.