Yeqa uye kokuyinhloko
Izifundo ze-AI

Yakha futhi Usebenzise I-ejenti Yocingo ye-AI Yasendaweni nge-ai-phone

Funda ukufaka i-ai-phone ku-Apple Silicon, ulungiselele amamodeli ayo enkulumo nolimi, usebenzise i-daemon yezingcingo, uyihlole endaweni nge-SIP, wenze izingcingo eziphumayo, futhi uhlanganise ama-ejenti angaphandle ngemodi ye-relay.

Yakha futhi Usebenzise I-ejenti Yocingo ye-AI Yasendaweni nge-ai-phone

Yini i-ai-phone?

ai-phone iyi-ejenti yocingo ye-AI esebenza kukhompyutha yakho. Ingamukela izingcingo ezingenayo, yenze izingcingo eziphumayo, iqhube izingxoxo ngokwemiyalelo ebhaliwe, futhi iveze imiyalo yesinyathelo ngesinyathelo ukuze ihlanganiswe ne-ejenti yangaphandle.

Inguqulo yamanje yakhelwe i-macOS ku-Apple Silicon. Ukuqashelwa kwenkulumo nokwenziwa kwenkulumo kusebenza endaweni nge-MLX, ngisho nalapho imodeli yolimi yengxoxo ikude. Ikhompyutha ephethe i-ai-phone kufanele kube i-Mac esekelwayo, kodwa abafonayo bangasebenzisa noma iyiphi ifoni noma iklayenti le-SIP.

Izici ezibalulekile

  • Izingcingo ezingenayo neziphumayo: Lungiselela uhlangothi olulodwa noma uvule zombili ngesikhathi sokusetha.
  • Ukucutshungulwa kwenkulumo endaweni: Ukuqashelwa kwenkulumo nokwenziwa kwezwi kuhlala ku-Mac.
  • Amamodeli olimi aguquguqukayo: Sebenzisa imodeli ye-MLX yasendaweni, okubhaliselwe kwe-Claude Code noma kwe-Codex, noma indawo yokugcina ehambisana ne-OpenAI.
  • Ukuxhumeka kwe-SIP nocingo: Hlola ngekheli le-SIP eliqondile noma uxhume i-SIP trunk, i-akhawunti ye-Telnyx, noma i-akhawunti ye-Twilio ukuze usebenzise izinombolo zocingo.
  • Ukuzenzakalela ngomugqa womyalo: Qala izingcingo ngomyalo bese uthola umbhalo wengxoxo ophumayo.
  • Imodi yokudlulisa: Vumela enye i-ejenti ekwazi ukusebenzisa i-shell ukuthi ilawule umzuliswano ngamunye wengxoxo.
  • Izinketho zezwi: Sebenzisa i-Phonon enezwi layo lesiNgisi elihlanganisiwe noma i-Kokoro enamazwi aseMelika nawaseBrithani kanye nesivinini esilungisekayo.

Njengamanje i-ai-phone isekela isiNgisi kanye nocingo olulodwa olusebenzayo ngesikhathi.

Izidingo

Inkundla esekelwayo

Udinga i-Mac ene-Apple Silicon. Ama-Mac e-Intel, i-Linux, ne-Windows akusekelwa okwamanje ngoba izinjini zangemuva zenkulumo zincike ku-MLX futhi ukufakwa kuhlanganisa ukuhlanganiswa okuqondene nenkundla.

Imemori nesitoreji

  • 8 GB RAM: Ukusetha kunikeza imodeli encane ye-2B.
  • 16 GB RAM: Kunconyelwa imodeli ezenzakalelayo ye-4B.
  • 32 GB RAM: Kunconyelwa imodeli ye-9B.
  • Isitoreji: Vumela amagigabhayithi ambalwa ukuze ulande izisindo zemodeli. Ukusetha kuhlola isikhala sediski esitholakalayo.

Isofthiwe nokuxhumeka kocingo

Faka i-Git, Homebrew, kanye ne-uv ngaphambi kokuqhubeka. Isifaki silawula amaphakheji e-Python adingekayo kanye nomtapo wolwazi we-SIP wendabuko.

Ezingcingweni, sebenzisa ikheli le-SIP ukuze uxhume ngokuqondile noma unikeze i-SIP trunk yakho, i-akhawunti ye-Telnyx, noma i-akhawunti ye-Twilio. Izindleko zomhlinzeki kanye nemikhawulo yokubhalisa kungase kusebenze.

Faka i-ai-phone

Yenza ikhophi yekhosombe, ungene kumkhombandlela walo, sebenzisa isifaki, bese uqala ukusetha okusebenzisanayo:

git clone https://github.com/oxnw/ai-phone.git
cd ai-phone
./scripts/install.sh
uv run ai-phone setup

Inqubo yokusetha ibuza ukuthi ufuna izingcingo ezingenayo, eziphumayo, noma zombili. Bese ilungiselela imodeli yolimi, ulayini wocingo, nezwi. Kumodeli yasendaweni, ukusetha kuncoma inketho ngokusekelwe kwimemori etholakele bese kulanda izisindo ezidingekayo.

Uma i-terminal yakho ingasebenzi kahle nemiyalo evamile yokusetha esebenzisanayo, sebenzisa imodi elula:

uv run ai-phone setup --plain

Qala i-daemon yocingo

Ngemva kokusetha, qala i-daemon usemkhombandleleni wekhosombe:

uv run ai-phone run

Gcina le nqubo isebenza noma nini lapho ufuna ukwamukela izingcingo. Uma usebenzisa izingcingo eziphumayo kuphela, yiqale ngaphambi kokwenza ucingo bese uyimisa ngo-Ctrl-C uma usuqedile.

Ukusetha akufaki isevisi esebenza ngemuva. Kufanele uziqalisele i-daemon ngemva kokusetha noma ngemva kokuqalisa kabusha ikhompyutha.

Yenza ucingo oluphumayo

Shiya i-daemon isebenza bese uvula enye i-terminal ekhosombeni. Sebenzisa i-call nenombolo oya kuyo kanye nomyalo ochaza umsebenzi:

uv run ai-phone call +15555550100 --json \
  --prompt "Buza ukuthi banalo yini itafula labantu ababili ngehora lesishiyagalombili namhlanje ebusuku."

Faka esikhundleni se-+15555550100 inombolo ohlose ukuyishayela. Lapho ucingo luphela, umyalo ubuyisa umbhalo wengxoxo. Lonke ucingo oluphumayo luqala ngesaziso sokuthi lusetshenziswa yi-AI.

Izingcingo ezingenayo zisebenzisa ukubingelela nemiyalelo egcinwe ngesikhathi sokusetha.

Yenza ukuhlola kwamahhala kwe-SIP yasendaweni

Ungahlola endaweni ngaphandle kokwenza ucingo lomhlinzeki. Qalisa ukusetha, uvule izingcingo ezingenayo, ukhethe A SIP address, bese ugcina igama elizenzakalelayo elithi agent. Nge-daemon ye-ai-phone esebenzayo, sebenzisa:

pjsua --local-port 5070 sip:agent@127.0.0.1

Lokhu kuxhuma iklayenti le-SIP lasendaweni ku-ejenti ekhelini le-loopback, okukuvumela ukuthi uqinisekise ukuhamba okuyisisekelo kocingo olungenayo ngaphambi kokulungiselela umhlinzeki wocingo.

Lawula izingcingo ngemodi yokudlulisa

Imodi yokudlulisa ihloselwe i-ejenti yangaphandle enokufinyelela ku-shell. Esikhundleni sokunikeza i-ai-phone umyalo owodwa wayo yonke ingxoxo, i-ejenti yangaphandle ingaqala ucingo, ikhulume, ilalele, bese inquma ukuthi yenzeni ngemva kwempendulo ngayinye.

Qala ucingo lokudlulisa

uv run ai-phone call +15555550100 --relay --json

Khuluma ulalele umzuliswano ngomzuliswano

uv run ai-phone say "Sawubona, ngicela ukubhukha itafula labantu ababili." --json
uv run ai-phone listen --json
uv run ai-phone say "Ngehora lesishiyagalombili, ngiyacela." --json

Qeda ingxoxo

uv run ai-phone hangup --say "Ngiyabonga. Sala kahle." --json

Umyalo ngamunye osebenzisa i---json ubhala into eyodwa ye-JSON kokuphumayo okujwayelekile. Umyalo we-call uphinde ubhale imicimbi yenqubekelaphambili emaphutheni ajwayelekile, ngakho ukuhlanganiswa kufanele kucubungule leyo misakazo ngokwehlukana.

Uma kungekho kholisho olusebenzayo, i-listen --json ilinda ikholi engenayo futhi iyiphathe ngemodi yokudlulisa. Ukujika kokudlulisa okungaphendulwanga kuphelelwa isikhathi ngemva kwemizuzwana engu-25 ngokuzenzakalelayo. Bheka ireferensi yemiyalo ye-agent ukuze uthole ama-ID amakholi, ukubuyiswa kombhalo okhulunywayo, izikhathi zokuphela ezilungisekayo, namakhodi okuphuma.

Khetha amamodeli namazwi

Ukuqaphela inkulumo

i-ai-phone isebenzisa i-Parakeet endaweni nge-MLX ukuze iqaphele inkulumo.

Imodeli yengxoxo

Imodeli yolimi ingaba imodeli ye-MLX yasendaweni, ukubhalisa kwakho kwe-Claude Code noma kwe-Codex, noma indawo yokugcina ehambisana ne-OpenAI. Ukusetha kukusiza ulungiselele inketho ekhethiwe.

Ukuhlanganiswa kwezwi

I-Phonon inikeza izwi lesiNgisi elihlanganisiwe. I-Kokoro inikeza amazwi aseMelika nawaseBrithani anesivinini sokukhuluma esilungisekayo.

Uma usebenzisa imodeli yolimi ekude, ithola umbhalo wekholi, imiyalelo, nomlando wengxoxo. Uma i-agent yangaphandle isebenzisa umhlinzeki wamafu, lowo mhlinzeki naye angathola noma yimiphi imibhalo ekhishiwe enikezwe yona. Ukuqaphela inkulumo nokuhlanganiswa kwezwi kusalokhu kwenzeka endaweni, kanti umsindo wefoni udlula ku-SIP naku-RTP.

Amathiphu okusetha okuthuthukile

Phinda usebenzise ukusetha lapho izidingo zishintsha

Qalisa ukusetha futhi ukuze ushintshe izwi, imodeli, izinkomba zamakholi ezinikwe amandla, noma i-ID yomshayi:

uv run ai-phone setup

Qalisa kabusha i-daemon ngemva kokushintsha ukumiswa ukuze izilungiselelo ezintsha zisebenze.

Qala nge-SIP yasendaweni ngaphambi kokwengeza inkampani yezokuxhumana

Qinisekisa i-agent ngokusebenzisa ukuhlolwa kwe-loopback SIP kuqala. Uma amakholi asendaweni esebenza, lungiselela i-SIP trunk, i-Telnyx, noma i-Twilio uma udinga izinombolo zocingo ezivamile. Ezinkingeni ezihlobene nenethiwekhi, buyekeza uhlu lokuhlola lwe-NAT lwephrojekthi.

Khetha imodeli ehambisana nenkumbulo etholakalayo

Landela isincomo esisekelwe ku-RAM sokusetha esikhundleni sokukhetha imodeli enkulu ngokuzenzakalelayo. Inketho engu-2B yenzelwe amasistimu anenkumbulo encane, kanti amamodeli azenzakalelayo angu-4B namakhulu angu-9B adinga inkumbulo eningi.

Cabangela ukubambezeleka kwamanje nesimo sokuhlolwa

I-README ibika impendulo yokuqala ezwakalayo emaphakathi engu-1.50 seconds ekuqalisweni kwasendaweni kwakamuva ngaphansi kwengcindezi enkulu yenkumbulo, okungafinyelelanga umgomo wephrojekthi wesekhondi elilodwa. Ukuhlolwa kwamakholi asendaweni nokuphazamiseka sekwenziwe, kodwa ukuhlolwa kokwamukelwa kwe-agent okubukhoma kwenkampani yezokuxhumana nokuphela kuya kokuphela kusalindile. Buyekeza izilinganiso ezishicilelwe kanye nomlando wokubambezeleka lapho uhlola ukulungela ukuqaliswa.

Sebenzisa amareferensi anemininingwane

Izinto okufanele zicatshangelwe ngelayisensi

i-ai-phone inelayisensi ngaphansi kwe-AGPL-3.0-or-later. Amafayela ayo esixhumi esibonakalayo se-PJSIP ahlanganisiwe asebenzisa i-GPLv2-or-later; bheka amanothi elayisensi ephrojekthi ngaphambi kokusabalalisa noma ukuhlanganisa le softhiwe.

Isiphetho

i-ai-phone inikeza indlela esebenzayo yomugqa womyalo yokusebenzisa i-agent yefoni ye-AI ku-Mac ene-Apple Silicon. Ngemva kokuyifaka, lungiselela imodeli nolayini we-SIP, qala i-daemon, bese uyihlola endaweni ngaphambi kokwenza amakholi enkampani yezokuxhumana. Ukuze wenze izinto ngokuzenzakalelayo, imodi yokudlulisa inikeza i-agent yangaphandle ukulawula okuhlelekile, ukujika ngakunye, ngemiyalo ye-shell ekhiqiza i-JSON.