Ming-3.0 Uyasebenza

I-Ming-3.0 iwuchungechunge lwamamodeli ezwi olukhishwe njenge-open source yi-Ant Group's Bailin, oluhlanganisa imisebenzi emithathu: ukuhlanganisa izwi, ukuqonda nomsindo nengxoxo, kanye nokubona izwi, ngamamodeli amahlanu. Uchungechunge lwe-Ming-3.0-tts lusekela “Imodi Yomqondisi” yokulawula ngokuningiliziwe imizwa, ijubane nokugcizelela; uchungechunge lwe-Ming-3.0-audio lungaqonda okurekhodiwe okude kusukela ekuqaleni kuze kube sekugcineni futhi lukhiphe imiphumela ehlelekile; uchungechunge lwe-Ming-3.0-asr lusekela ukufakwa kwamagama abalulekile nokubona izilimi zesigodi. Amamodeli asetshenziswa emisebenzini eminingi yangaphakathi ye-Ant Group, futhi kunikezwa ulwazi oluku-inthanethi kanye nezinsizakalo ze-API.

Voice Design ukwakhiwa kombala wezwi: yenza ngokwezifiso umbala wezwi owufunayo ngokuwuchaza ngombhalo noma usebenzise umsindo oyisibonelo.
Imodi Yomqondisi: lawula ngokuningiliziwe izinguquko zemizwa, izindawo zokugcizelela kanye nokuphathwa kwejubane lezwi usebenzisa imiyalelo.
Amathegi ezici ezingezona izwi emgqeni: faka izimpawu ezifana nokuhleka nokububula embhalweni ukuze isimo somlingiswa siphile kakhudlwana.
Ukuqonda umsindo omude kusukela ekuqaleni kuze kube sekugcineni: bhala phansi, fingqa futhi ukhiphe ukuvumelana nokungavumelani okurekhodiwe okude ngokuqondile, kugcinwe imizwa nolwazi lwabakhulumayo.
Ukucabanga ngezwi nokubiza amathuluzi: sebenzisa ukuqonda okurekhodiwe ukucabanga nokubiza amathuluzi angaphandle, kusekelwe imisebenzi yezwi ephelele.
Ukufakwa kwamagama abalulekile ku-Context ASR: dlulisa uhlu lwamagama abalulekile nomongo wengxoxo ngesikhathi sokuqaphela ukuze kubonwe ngokunembile izindlela zokuphimisa, amagama ezindawo namagama ochwepheshe.
Ukuhamba komsebenzi okuhlanganisa amamodeli: amakhono amathathu e-TTS, Audio ne-ASR angahlanganiswa ngokukhululeka ukuze axhunywe ekuhambeni komsebenzi konjiniyela.

Landela i-WeChat bese uphendula ngokuthi “开源” ukuze ujoyine iqembu lengxoxo lamaphrojekthi e-AI open source
Uchungechunge lwe-Audio: isakhiwo sokuqonda umsindo omude. I-Ming-3.0-audio-flash isebenzisa ukumelwa komsindo okunezinga eliphansi lamafreyimu elingu-6.25 Hz ohlangothini lokufaka, inciphise ubude bokulandelana obuhambisana nokurekhodiwe okufanayo; isisekelo solimi sisekelwe ku-Ling-3.0, sihlanganisa ukunaka okuxubile kwe-KDA ne-Gated MLA ukuze kubanjwe ukuncika kwesikhathi eside ngesikhathi kulawulwa inani lokubala. Bese kwengezwa isakhiwo se-sparse MoE ukuze kuncishiswe nakakhulu umthwalo wokucubungula, okwenza kube nokwenzeka ukubuza ngokuqhubekayo nokuphendula ngokushesha mayelana nokurekhodiwe okude, kugcinwe imininingwane eningiliziwe efana nezwi lomuntu, imicimbi yomsindo nomsindo wemvelo.
Uchungechunge lwe-TTS: ukukhiqizwa kwezwi okulawulwa yizimo. I-Ming-3.0-tts ifinyelela ukuhlanganiswa kwemvelo kakhulu ngosayizi omncane wamapharamitha ongu-0.5B. Umqondo oyinhloko uwukusebenzisa imiyalelo yomqondisi kanye nencazelo yombala wezwi/umsindo oyisibonelo njengokufakwayo okulawulayo, kuhlanganiswe nombhalo wokuhlanganisa ukuze kulawulwe ukukhiqizwa kwe-acoustic, okwenza imodeli iveze amazinga obuciko anobuciko ngencazelo yombhalo.
Amathegi ezici ezingezona izwi e-TTS: ukulawula imicimbi ye-acoustic ngokusobala. Ngokufaka amathegi asemgqeni afana no-[Laughter] embhalweni, indawo yomcimbi wezwi ongajwayelekile yaziswa imodeli ngokusobala, okwenza kube nokulawula okunembile kokufakwa kokuhleka, ukububula, ukukhwehlela nezinye izinkulumo.
Uchungechunge lwe-ASR: ubunjiniyela bekhophasi + ukufakwa komongo. I-Ming-3.0-asr-tiny yakha ikhophasi esezingeni eliphezulu ehlanganisa isiMandarin, isiNgisi, izilimi zesigodi namagama abalulekile emikhakheni ethile, ilawule ngokufanele ukusatshalaliswa kwedatha. Ngesikhathi esifanayo, isebenzisa ikhono lokuqonda umongo lesisekelo semodeli yolimi ukufaka uhlu lwamagama abalulekile angaba khona kanye nomongo wengxoxo njengolwazi lwezimo enqubweni yokuqaphela, kusize ekwahlukaniseni amagama anomsindo ofanayo namagama angavamile, ngaphandle kokuqeqeshwa okwengeziwe kwezimo ezithile.

Ulwazi oluku-inthanethi: vakashela ku-https://ling.tbox.cn, bese ukhetha i-Ming-3.0-audio-flash ku-Ling Studio Chat. Layisha umsindo ukuze uzame amakhono okuqonda umsindo nengxoxo.
Ukuxhuma i-API: sebenzisa i-Ling Studio ukubiza izinsizakalo ze-API ezivulekele amamodeli amathathu, i-tts-0.5B, i-audio-flash ne-asr-tiny, ukuze uhlanganise ukubona izwi nokukhiqiza izwi ekuhambeni komsebenzi wakho.
Ukufaka endaweni: landa izisindo zamamodeli ku-Hugging Face, bese uzilayisha futhi uzisebenzise endaweni yangakini.
Ukuhlanganisa ekuhambeni komsebenzi: hlanganisa amakhono ahlukene e-TTS, Audio ne-ASR ngokwezidingo ukuze wenze umjikelezo ophelele ofana nokuthi “ukufakwa kwezwi→ukuqonda nokuhlaziya→ukusakazwa kwezwi”.

Imisebenzi emithathu ihlanganiswe ngokuphelele: imigqa emithathu yemikhiqizo namamodeli amahlanu okuhlanganisa izwi, ukuqonda umsindo nokubona izwi, okwenza onjiniyela bakhethe noma bahlanganise amakhono ngokwezidingo ekuhambeni komsebenzi.
Imodi Yomqondisi iyinto entsha embonini: imiyalelo yombhalo kuphela yenza kube nokulawula okunembile kwemizwa, ukugcizelela nejubane, kukhiqizwe izwi elisezingeni lokulingisa obekunzima ukulifinyelela nge-TTS yendabuko.
0.5B wamapharamitha amancane ufinyelela i-SOTA: ukuqina kokuphimisa okungu-93.2% kubekwe endaweni yokuqala ngokulingana kuwo wonke amazinga, kanti i-Controlled Voice ene-1069 Elo ibekwe endaweni yokuqala kumamodeli anesisindo esivulekile. Izindleko zokuyifaka ziphansi kakhulu kunamamodeli amakhulu asezingeni elifanayo.
Ukuqonda umsindo omude kusukela ekuqaleni kuze kube sekugcineni: kugcina ngokuqondile ulwazi olufana nemizwa, ubunikazi bomkhulumi nomsindo wemvelo, kwenze ukulungiswa kwamaminithi emihlangano nezingxoxo kube umsebenzi owodwa.
Ukucabanga okusebenzayo ngomsindo omude: ukumelwa kwezinga eliphansi lamafreyimu elingu-6.25Hz + ukunaka okuxubile kwe-KDA/Gated MLA + sparse MoE kunciphisa kakhulu umthwalo wokubala wokuqonda nokuxhumana nokurekhodiwe okude.
Ukufakwa kwamagama abalulekile nomongo ngaphandle kokuqeqeshwa: ngesikhathi sokubona izwi nge-ASR, uhlu lwamagama abalulekile nomongo wengxoxo kungadluliswa ngokuqondile; ukubona izilimi zesigodi, amagama ezindawo namagama ochwepheshe akudingi ukuqeqeshwa okwengeziwe.

Iqoqo lamamodeli le-HuggingFace: https://huggingface.co/collections/inclusionAI/ming-30

Ukuhlelwa kwamaminithi emihlangano nezingxoxo: sebenzisa i-audio-flash ukucubungula okurekhodiwe okude ngokuqondile, ukhiphe ngokuzenzakalelayo “izixazululo okuvunyelwene ngazo, izindaba ezisalindele isinqumo, nezizathu zezinhlangothi ezahlukene”, bese ukhipha imiphumela ehlelekile njengamaminithi omsebenzi, kuncishiswe isikhathi sokubhala umbhalo wezwi ngesandla.
Okuqukethwe okuzwakalayo nokudlalwa kwabalingiswa: amaqembu okuqukethwe asebenzisa i-Voice Design ne-Imodi Yomqondisi ye-TTS ukukhiqiza izwi elihambisana ngokuphelele nemizwa, ijubane nesitayela semigqa ngokwezici zomlingiswa, futhi angafaka ukuhleka nokububula ukuze umlingiswa aphile kakhudlwana.
Ukuguqula izwi lamakhasimende nawemisebenzi libe umbhalo: i-asr-tiny ifaka amagama abalulekile embonini, njengamagama emikhiqizo namagama ezindawo, kanye nomongo wengxoxo, ukuze ibone ngokunembile izwi lamakhasimende elinokuphimisa kwesigodi. Ilungele izimo ezifana namarekhodi esevisi yamakhasimende, ukusesha kwe-Alipay nokuqinisekiswa kobunikazi ngokuphepha.
Abasizi bezwi bemikhiqizo: hlanganisa i-TTS ne-ASR ukuze uxhume izinyathelo ezimbili, ukufakwa kwezwi→ukuqonda incazelo→ukusakazwa kwezwi, njengengxoxo eyinhloko nezingcingo zezwi kuhlelo lokusebenza i-Afu, kanye nomsizi wokukopisha umbala wezwi wodokotela.