Hy-SimulMT Uyasebenza

I-Hy-SimulMT iyimodeli enkulu yokuhumusha ngesikhathi esisodwa ekhishwe yi-Tencent Hunyuan njengomthombo ovulekile. Isekela ukuhumusha kwesikhathi sangempela kusuka enkulumweni kuya enkulumweni phakathi kwesiShayina nesiNgisi, ikwazi ukuhumusha ngesikhathi ulalele futhi ikopishe ngokushesha iphimbo lesikhulumi. Ukuhlolwa kwe-Hy-SimulMT kubonisa ukubambezeleka okungaba imizuzwana engu-3, okungcono kunezinga labahumushi abangabantu elingu-4–5 imizuzwana; kusethi yokuhlola ye-RealSI, inokubambezeleka okuphansi kakhulu kanye nekhwalithi yokuhumusha ehamba phambili phakathi kwamamodeli omthombo ovulekile, idlula ama-API amaningi ezentengiselwano, futhi ekuhlolweni emikhakheni eminingi ikhwalithi yayo idlula ochwepheshe abangabantu bokuhumusha ngesikhathi esisodwa.

Ukuhumusha kwenkulumo ngesikhathi sangempela: Umsindo wesikhulumi ungangena ngokuqhubekayo kumodeli, imodeli ilalele futhi ihumushe ngesikhathi esifanayo, ikhiphe umbhalo ohunyushiwe njengoba ikhuluma, ngaleyo ndlela yenze ukuhumusha kwesikhathi sangempela phakathi kwesiShayina nesiNgisi kusuka enkulumweni kuya enkulumweni.
Ukukopishwa kwephimbo ngesikhathi sangempela: Okukhishwayo okuhunyushiwe kungagcina iphimbo nendlela yokukhuluma yesikhulumi uqobo; ezingxoxweni zabantu abaningi ingahlukanisa futhi ibuyisele amazwi ezikhulumi ezahlukene.
Ukukhishwa kombhalo nenkulumo: Imodeli ingakhipha umbhalo ohunyushiwe (inguqulo ye-Interleave ingakhipha nombhalo wolimi oluwumthombo ngesikhathi esifanayo), kanti isevisi ye-TTS ehambisanayo ingakhipha inkulumo ehunyushiwe.
Ukusabela okunokubambezeleka okuphansi: Ukuhlolwa kubonisa ukubambezeleka kokuhumusha ngesikhathi esisodwa okungaba imizuzwana engu-3, okushesha kunezinga labahumushi abangochwepheshe elingu-4.5–4.7 imizuzwana.

Landela i-WeChat bese uphendula ngokuthi “开源” ukuze ujoyine iqembu lokuxhumana lamaphrojekthi e-AI omthombo ovulekile
Ukuhumusha okuqhubekayo ngesikhathi esisodwa: ukuqonda okwandayo kanye nokuthi “qala ukukhuluma, ulungise kamuva”. Ukungqubuzana okuyinhloko ekuhumusheni ngesikhathi esisodwa kusekulinganiseni phakathi kokuthi “ulinde ulwazi olwengeziwe ukuze uhumushe ngokunembile” kanye nokuthi “uqale ukukhuluma kusenesikhathi ukuze ukulalela kube kuhle”. I-Hy-SimulMT ayilindi ukuthi isikhulumi siqede ukukhuluma; ithola futhi iqonde kancane kancane emsindweni oqhubekayo, iqoqe ulwazi oluncane lwenkulumo bese iqala ukukhiqiza ukuhumusha. Uma ulwazi olulandelayo lufika futhi kutholakala ukuthi okukhiphwe ngaphambilini akuphelele noma kunephutha, ilungisa ukuhumusha ngendlela efana nokuzilungisa komhumushi ongumuntu.
Injini enkulu yokuhumusha ye-MoE enokusebenza okungajwayelekile: Imodeli yomthombo ovulekile yi-Hy-SimulMT-30B-A3B, okuwuhlaka lwe-Mixture of Experts (MoE) olunamapharamitha angu-30B esewonke, kanti cishe amapharamitha angu-3B ayasebenza ngesikhathi ngasinye. Isekelwe emndenini wamamodeli we-Hunyuan HYV3ForCausalLM futhi isebenzisa i-vLLM njengengemuva eliyinhloko lokuqagela. Ukwakhiwa kokusebenza okungajwayelekile kuyenza igcine ikhwalithi yokuhumusha yemodeli enkulu kuyilapho izindleko zokuqagela zisondela kwezemodeli encane, okuyenza ifanele izimo ze-inthanethi zokuhumusha ngesikhathi esisodwa ezizwela isivinini.
Ukukopishwa kwephimbo kwenziwa ingxenye ezimele ye-TTS: Umzimba wemodeli uphethe ukuqaphela inkulumo okuqhubekayo, ukuhumusha ngesikhathi esisodwa, nokukhipha umbhalo (kunezindlela ezimbili zokukhipha: ukuhumusha kuphela, noma umbhalo womthombo kanye nokuhumusha), kanti ukuhlanganiswa kwenkulumo nokukopishwa kwephimbo lokufaka kuphathwa isevisi ye-TTS elungisekayo. Lokhu kwakhiwa kwamamojuli kwenza ikhwalithi yokuhumusha nobunjalo bephimbo kuthuthukiswe ngokwehlukana futhi kwenze kube lula ukuxhuma i-TTS yezinkampani zangaphandle; kodwa amaphutha okuqaphela e-ASR ochungechungeni oluxhumene angadlulela ezingxenyeni ezilandelayo, ngakho kudingeka ukubhekwana nawo ngokuqina kwemodeli yokuhumusha.

Clone inqolobane: sebenzisa git clone https://github.com/Tencent-Hunyuan/Hy-SimulMT.git bese ungena enkombeni yephrojekthi.
Faka isevisi yokuqagela ngomyalo owodwa: ngena enkombeni ye-SimulInfer/ bese usebenzisa bash deploy.sh --install-server --model tencent/Hy-SimulMT-30B-A3B-Interleave.
Lungisa i-TTS (ongakukhetha): Uma udinga okukhiphayo kwenkulumo nokukopishwa kwephimbo, lungisa ukuxhumeka kwesevisi ye-TTS ku-config.toml; ukuhumusha kombhalo kuphela akudingi lesi sinyathelo.
Qala ukusebenzisa: linda kuze kubonakale Ready: http://127.0.0.1:8080/ esigungwini, vula lelo kheli esipheqululini, khetha indlela yokuhumusha, bese ulayisha ifayela lomsindo noma uvule imakrofoni ukhulume. Chofoza u-「Stop and finish」 ukuze uqede.
Ukubiza ngomugqa womyalo: sebenzisa curl -N http://127.0.0.1:8080/translate/stream -F direction=en-zh -F file=@audio.wav ukuze ulayishe ifayela, noma usebenzise iklayenti le-simulinfer-client eliza nenqolobane ukuze uthole imiphumela yokuhumusha.

Ukuhola ngokubambezeleka okuphansi: I-StreamLAAL ehloliwe ingaba imizuzwana engu-3, ishesha kunomhumushi onguchwepheshe futhi iyimodeli yomthombo ovulekile yokuhumusha ngesikhathi esisodwa enokubambezeleka okuphansi kakhulu phakathi kwezinhlelo eziqhathanisiwe.
Ikhwalithi nesivinini ndawonye: Ekuhlolweni komsindo omude kwe-RealSI, ifinyelela “ekubambezelekeni okuphansi kakhulu + kwikhwalithi ephezulu kakhulu”; amaphuzu e-GEMBA adlula wonke amamodeli omthombo ovulekile aqhathanisiwe nama-API ezentengiselwano angu-4 kwangu-5.
Ifinyelela ezingeni lochwepheshe abangabantu: Ekuhlolweni okuyimpumputhe okuqhathanisa ochwepheshe abangu-6 bokuhumusha ngesikhathi esisodwa abanezitifiketi ze-CATTI ezingeni lokuqala/lesibili, ithole amaphuzu aphezulu ekhwalithi yokuhumusha futhi yasabela ngokushesha.
Ukukopishwa kwephimbo ngesikhathi sangempela: Inkulumo ehunyushiwe ingabuyisela kakhulu iphimbo nendlela yokukhuluma yesikhulumi; ezimweni zabantu abaningi ingahlukanisa izikhulumi ezahlukene, yenze ukuxhumana kube ngokwemvelo.

Inqolobane ye-GitHub: https://github.com/Tencent-Hunyuan/Hy-SimulMT
Inqolobane yamamodeli ye-HuggingFace: https://huggingface.co/collections/tencent/hy-simulmt

Imihlangano nezinkundla zamazwe ngamazwe: Imicimbi emikhulu yamazwe ngamazwe iyikhetha njengohlelo olusemthethweni lokuhumusha lwe-AI lwamavolontiya nabasebenzi bezinsizakalo basezingeni eliphansi, olusetshenziselwa ukwamukela izivakashi zakwamanye amazwe nokuxhumana esizeni.
Izingxoxo nemihlangano yebhizinisi yezilimi ezahlukene: Emihlanganweni yezinkampani zamazwe amaningi, yenza ukuhumusha kwesikhathi sangempela phakathi kwesiShayina nesiNgisi, ukuze bonke ababambiqhaza bakhulume futhi balalele ngolimi lwabo lwendabuko, kuncishiswe izindleko zokuxhumana.
Ukuhumusha ngesikhathi esisodwa ezinkulumweni nasekuqeqeshweni: Ezinkulumweni zemfundo, ekuqeqeshweni kwangaphakathi ezinkampanini, ezifundweni ezivulekile nakwezinye izimo lapho umuntu oyedwa ekhuluma isikhathi eside, imodeli ingalandela futhi ihumushe ngokuzinzile, ngenkulumo ehunyushiwe yemvelo, ifaneleka njengethuluzi “lokusiza ukuqonda”.
Izingxoxo zabantu abaningi nokuqukethwe kwabezindaba: Ezingxoxweni zabantu abaningi njengengxoxo, ama-podcast namadokhumentari, ukuhumusha ngesikhathi esisodwa kukopisha iphimbo lesikhulumi ngasinye, ukuze abalaleli bakwazi ukuhlukanisa ukuthi ubani okhulumayo.
Ezokuvakasha nezinsizakalo zansuku zonke ezibhekene nabantu bakwamanye amazwe: Ezimweni zokuxhumana ezinobungozi obuphansi ezindaweni zezinsizakalo ezifana nezikhumulo zezindiza, amahhotela, imibukiso nezindawo zokudlela, kusiza abasebenzi bezinsizakalo nezivakashi zakwamanye amazwe ukuthi baxhumane ubuso nobuso ngendlela eyisisekelo esebenzisekayo.