I-Nvidia PersonaPlex 7B ku-Apple Silicon: I-Full-Duplex Speech-to-Speech ku-Swift

Sethula i-Frontier Entsha ye-Voice AI

Ukwakheka kwezwe kobuhlakani bokwenziwa kuyashintsha kusuka emafini kuya onqenqemeni, futhi i-Apple Silicon ihola phambili. Konjiniyela, amandla okusebenzisa amamodeli anamandla endaweni avula umhlaba omusha wamathuba wezinhlelo zokusebenza ezisabelayo, eziyimfihlo, nezingakwazi ukuxhuma ku-inthanethi. Faka i-PersonaPlex 7B ye-Nvidia, imodeli yesimanje eyenzelwe i-AI yengxoxo yemvelo, ezwakalayo. Uma le modeli enamandla ibhangqwe namandla enjini ye-neural ye-M-series Mac kanye nokusebenzisa okulula kwe-Swift, umphumela uba impumelelo esikhathini sangempela, ukusebenzisana kwenkulumo-kuya-enkulumweni okuphindwe kabili okugcwele.

Iyini Inkulumo-kuya-Inkulumo Egcwele Ekabili?

Ngaphambi kokuthi ungene kumlingo wobuchwepheshe, kubalulekile ukuqonda ingxenye ye-"full-duplex". Ngokungafani nabasizi bezwi abalula abadinga ukuthi ucindezele inkinobho bese ulinda impendulo, ukusebenzisana kwe-full-duplex kulingisa ingxoxo yomuntu yemvelo. Ivumela ukukhuluma nokulalela kanyekanye, ivumele ukuphazamiseka, ukumisa isikhashana, kanye nengxoxo yangempela yemuva naphambili. Lokhu kusho ukuthi i-AI ingacubungula okushoyo ngenkathi usakhuluma futhi yakhe impendulo eqala lapho uqeda—noma ingene kancane uma umisa isikhashana. Ukufeza lokhu kudivayisi yasendaweni, ngaphandle kokuthumela umsindo kuseva ekude, kuwumsebenzi ongcwele wokudala ulwazi lomsebenzisi olungenazihibe futhi olunembile.

Isebenzisa I-Apple Silicon's Unified Architecture

Ukhiye wokwenza lokhu kwenzeke kukhompuyutha ephathekayo noma ideskithophu isakhiwo esiyingqayizivele se-Apple Silicon. Ama-chips ochungechunge lwe-M ahlanganisa i-CPU, i-GPU, ne-Neural Engine enamandla (NE) ocezwini olulodwa lwe-silicon. Lesi sakhiwo senkumbulo esihlanganisiwe silungele imithwalo yokufunda ngomshini. Amamodeli amakhulu afana ne-PersonaPlex 7B angalayishwa ngokuqondile kumemori eyabiwe, okuvumela i-CPU ukuthi isingathe i-logic yohlelo lokusebenza nge-Swift, i-GPU ukuthi isheshise ukubala okuthile, kanye ne-Neural Engine ukuthi idabule ukusebenza kwe-core tensor yemodeli ngendlela esebenza kahle kakhulu. Lokhu kusebenzisana kususa izingqinamba zokuhambisa idatha phakathi kwezingxenye ezihlukene, okwenza ukuqagela kwesikhathi sangempela kungenzeki nje, kodwa kushelele futhi konga amandla.

• Ubumfihlo Nesivinini: Konke ukucubungula kwenzeka endaweni kudivayisi. Izingxoxo zakho ezibucayi azithunyelwa emafini, okuqinisekisa ubumfihlo bedatha obuphelele kuyilapho uzuza ngokubambezeleka okucishe kube yiziro.
• Ukusebenza Okungaxhunyiwe ku-inthanethi: Izinhlelo zokusebenza ezakhiwe ngalesi sitaki zisebenza noma kuphi, ngaphandle koxhumano lwe-inthanethi, okuzenza zithembeke ngendlela emangalisayo.
• Ukusebenza Komdabu: Ukusebenzisa i-Swift nezinhlaka zomdabu ezifana ne-Core ML kuvumela ukuhlanganiswa okujulile ne-macOS, okuholela ekuzizweleni okubushelelezi okuzwakala kuyingxenye yesistimu yokusebenza ngokwayo.

Ukwakha Ipayipi nge-Swift

Ukudala leli payipi le-duplex egcwele ku-Swift kuhlanganisa ukuhlela izingxenye ezimbalwa. Okokuqala, uhlaka lwe-AVFoundation luthwebula okokufaka komsindo kusuka kumakrofoni. Lokhu kusakaza komsindo kuguqulwa kube umbhalo kusetshenziswa imodeli yendawo yokuqaphela inkulumo, njengohlaka lwe-Apple olukudivayisi Lwenkulumo. Umbhalo owumphumela ufakwa kumodeli ye-Nvidia PersonaPlex 7B, elungiselelwe ukuthi isebenze nge-Core ML noma enye injini ye-inference ehambisana ne-Swift efana ne-MLX. Imodeli ikhiqiza impendulo yombhalo ecabangelayo, eqaphela umongo. Ekugcineni, lo mbhalo uguqulelwa emuva ube yinkulumo efana nempilo kusetshenziswa injini yendawo yokuguqula umbhalo ube inkulumo (TTS). Inselele yangempela isekuphatheni lezi zingxenye kanye kanye ukuze kuzuzwe umphumela we-duplex egcwele—umsebenzi lapho imodeli yesimanje ye-Swift yokuvumelana ne-async/await ihamba phambili.

"Amandla okusebenzisa imodeli yaleli zinga endaweni ku-Apple Silicon ashintsha kakhulu indlela esicabanga ngayo ngokuhlanganisa i-AI ekusebenzeni kwethu kwansuku zonke. Isusa i-AI isuka kusevisi exhunyiwe iye kuthuluzi lomdabu, elihlala litholakala." – Unjiniyela Omkhulu e-Mewayz

Imithelela Yezingxenyekazi Ezifana Ne-Mewayz

Kusistimu yokusebenza yebhizinisi efana ne-Mewayz, lokhu kugxuma kobuchwepheshe kuyashintsha. Cabanga ngabenzeli bezwi abahlakaniphile ngaphakathi kwesofthiwe yebhizinisi lakho abangakusiza ukuthi ubhale ama-imeyili, uphathe imigqa yesikhathi eyinkimbinkimbi yephrojekthi, noma uhlaziye idatha—konke ngengxoxo yemvelo, ngaphandle kokonakalisa idatha yenkampani ebucayi. Imojula ye-Mewayz enikwa amandla yi-PersonaPlex 7B yendawo inganikeza:

Ubumfihlo obungakaze bubonwe bokuphatha ulwazi lwebhizinisi oluyimfihlo, izinsiza zokusebenzisana zesikhathi sangempela ezizwakala sengathi unozakwenu onguchwepheshe egumbini, kanye nama-bot osekelo lwamakhasimende asabela kakhulu ahlanganiswe ngokuqondile endaweni yokusebenza yenkampani ye-Mewayz. Lobu buchwepheshe buvumela i-Mewayz ukuthi idlule imingcele yalokho i-OS yebhizinisi engakwazi ukuyenza, idlulele ngale kokuxhumana okulula okusekelwe emiyalweni iye kumasistimu ezingxoxo ngempela nezihlakaniphile akhulisa ikhono lomuntu ngokuphepha nangempumelelo.

Inhlanganisela ye-Nvidia's PersonaPlex 7B, amandla aluhlaza e-Apple Silicon, nobuhle be-Swift ingaphezu nje kwedemo yobuchwepheshe; ipulani yekusasa lokusebenzisana kwabantu namakhompyutha. Ithembisa umhlaba lapho izisetshenziswa zethu zingasiqondi njengabakhuzi abanikeza ama-oda, kodwa njengozakwethu engxoxweni.

Imibuzo Evame Ukubuzwa

Sethula i-Frontier Entsha ye-Voice AI

Ukwakheka kwezwe kobuhlakani bokwenziwa kuyashintsha kusuka emafini kuya onqenqemeni, futhi i-Apple Silicon ihola phambili. Konjiniyela, amandla okusebenzisa amamodeli anamandla endaweni avula umhlaba omusha wamathuba wezinhlelo zokusebenza ezisabelayo, eziyimfihlo, nezingakwazi ukuxhuma ku-inthanethi. Faka i-PersonaPlex 7B ye-Nvidia, imodeli yesimanje eyenzelwe i-AI yengxoxo yemvelo, ezwakalayo. Uma le modeli enamandla ibhangqwe namandla enjini ye-neural ye-M-series Mac kanye nokusebenzisa okulula kwe-Swift, umphumela uba impumelelo esikhathini sangempela, ukusebenzisana kwenkulumo-kuya-enkulumweni okuphindwe kabili okugcwele.

Iyini Inkulumo-kuya-Inkulumo Egcwele Ekabili?

Ngaphambi kokuthi ungene kumlingo wobuchwepheshe, kubalulekile ukuqonda ingxenye ye-"full-duplex". Ngokungafani nabasizi bezwi abalula abadinga ukuthi ucindezele inkinobho bese ulinda impendulo, ukusebenzisana kwe-full-duplex kulingisa ingxoxo yomuntu yemvelo. Ivumela ukukhuluma nokulalela kanyekanye, ivumele ukuphazamiseka, ukumisa isikhashana, kanye nengxoxo yangempela yemuva naphambili. Lokhu kusho ukuthi i-AI ingacubungula okushoyo ngenkathi usakhuluma futhi yakhe impendulo eqala lapho uqeda—noma ingene kancane uma umisa isikhashana. Ukufeza lokhu kudivayisi yasendaweni, ngaphandle kokuthumela umsindo kuseva ekude, kuwumsebenzi ongcwele wokudala ulwazi lomsebenzisi olungenazihibe futhi olunembile.

Isebenzisa I-Apple Silicon's Unified Architecture

Ukhiye wokwenza lokhu kwenzeke kukhompuyutha ephathekayo noma ideskithophu isakhiwo esiyingqayizivele se-Apple Silicon. Ama-chips ochungechunge lwe-M ahlanganisa i-CPU, i-GPU, ne-Neural Engine enamandla (NE) ocezwini olulodwa lwe-silicon. Lesi sakhiwo senkumbulo esihlanganisiwe silungele imithwalo yokufunda ngomshini. Amamodeli amakhulu afana ne-PersonaPlex 7B angalayishwa ngokuqondile kumemori eyabiwe, okuvumela i-CPU ukuthi isingathe i-logic yohlelo lokusebenza nge-Swift, i-GPU ukuthi isheshise ukubala okuthile, kanye ne-Neural Engine ukuthi idabule ukusebenza kwe-core tensor yemodeli ngendlela esebenza kahle kakhulu. Lokhu kusebenzisana kususa izingqinamba zokuhambisa idatha phakathi kwezingxenye ezihlukene, okwenza ukuqagela kwesikhathi sangempela kungenzeki nje, kodwa kushelele futhi konga amandla.

Ukwakha Ipayipi nge-Swift

Ukudala leli payipi le-duplex egcwele ku-Swift kuhlanganisa ukuhlela izingxenye ezimbalwa. Okokuqala, uhlaka lwe-AVFoundation luthwebula okokufaka komsindo kusuka kumakrofoni. Lokhu kusakaza komsindo kuguqulwa kube umbhalo kusetshenziswa imodeli yendawo yokuqaphela inkulumo, njengohlaka lwe-Apple olukudivayisi Lwenkulumo. Umbhalo owumphumela ufakwa kumodeli ye-Nvidia PersonaPlex 7B, elungiselelwe ukuthi isebenze nge-Core ML noma enye injini ye-inference ehambisana ne-Swift efana ne-MLX. Imodeli ikhiqiza impendulo yombhalo ecabangelayo, eqaphela umongo. Ekugcineni, lo mbhalo uguqulelwa emuva ube yinkulumo efana nempilo kusetshenziswa injini yendawo yokuguqula umbhalo ube inkulumo (TTS). Inselele yangempela isekuphatheni lezi zingxenye kanyekanye ukuze kuzuzwe umphumela we-duplex egcwele—umsebenzi lapho imodeli yesimanje ye-Swift yokuvumelana ne-async/await iphumelela kahle kakhulu.

Imithelela Yezingxenyekazi Ezifana Ne-Mewayz

Kusistimu yokusebenza yebhizinisi efana ne-Mewayz, lokhu kugxuma kobuchwepheshe kuyashintsha. Cabanga ngabenzeli bezwi abahlakaniphile ngaphakathi kwesofthiwe yebhizinisi lakho abangakusiza ukuthi ubhale ama-imeyili, uphathe imigqa yesikhathi eyinkimbinkimbi yephrojekthi, noma uhlaziye idatha—konke ngengxoxo yemvelo, ngaphandle kokonakalisa idatha yenkampani ebucayi. Imojula ye-Mewayz enikwa amandla yi-PersonaPlex 7B yendawo inganikeza: