Yeqa uye kokuyinhloko
Izifundo ze-AI

Yakha Isitudiyo Sendawo Somculo Nemisindo Ekhiqizwa Yi-AI nge-Audio Studio MCP

Lesi sifundo sichaza indlela yokufaka i-Audio Studio MCP kuma-Mac ane-Apple Silicon noma kumasistimu e-Windows ane-NVIDIA, ukuyixhuma kuklayenti le-MCP, ukukhiqiza umculo nemisindo nge-Stable Audio 3, nokucwengisisa imiphumela ngamathuluzi okuhlela, ukuphinda, ukulinganisa umsindo nokuthekelisa.

Yakha Isitudiyo Somculo Nemisindo Ye-AI nge-Audio Studio MCP

Audio Studio MCP iguqula ikhompyutha yendawo ye-Mac noma ye-Windows ibe isitudiyo somsindo esinikwa amandla yi-AI se-Claude Code, Claude Desktop, Cursor, namanye amaklayenti e-Model Context Protocol. Uchaza ngolimi olujwayelekile umsindo owudingayo, bese iseva isebenzisa i-Stable Audio 3 endaweni ukuze ikhiqize umculo, izihibe zezinsimbi, noma imiphumela yomsindo.

Ngenxa yokuthi ukucubungula kwenzeka emshinini wakho, iphrojekthi ayidingi isevisi yamafu noma ukhiye we-API. Amafayela akhiqiziwe agcinwa endaweni njengamafayela e-WAV ane-sidecar ye-JSON equkethe i-prompt, i-seed, nezilungiselelo ezisetshenzisiwe ukuwakha. Ngemva kwalokho ungahlela, wenze iluphu, wenze i-normalize, futhi uthumele imiphumela ku-game engine noma ku-video editor.

Ukuhamba komsebenzi kokukhiqizwa komsindo wendawo nge-Audio Studio MCP

Okungenziwa yi-Audio Studio MCP

Iphrojekthi inikeza iseva ye-MCP phakathi kweklayenti lakho le-AI ne-runtime yendawo ye-Stable Audio 3. Isicelo esinjengokuthi “Yenza iluphu ye-lo-fi BGM yemizuzwana engu-60 ku-85 BPM yesehlakalo sekhefi” sithunyelwa nge-MCP, siguqulwe sibe yi-prompt ehambisana ne-Stable Audio, bese sicutshungulwa yi-backend yendawo esekelwayo.

Kuma-Apple Silicon, i-Audio Studio MCP isebenzisa i-backend ye-MLX. Ku-Windows ene-NVIDIA GPU, isebenzisa i-backend ye-PyTorch ene-CUDA. Okuphumayo kungagcinwa njengefayela le-WAV lekhwalithi ephezulu noma kucutshungulwe kube yi-OGG ne-FLAC.

Amakhono abalulekile

  • Khiqiza umculo, izihibe zezinsimbi, ama-stem, nemiphumela yomsindo kusuka embhalweni.
  • Cela amakhandidethi aphakathi kuka-1 no-4 ocingweni olulodwa.
  • Shintsha isitayela se-clip yonke, ukhiqize kabusha ububanzi besikhathi obukhethiwe, noma welule i-track ekhona.
  • Dala izihibe ezingenamthungo usebenzisa ama-crossfade.
  • Susa ukuthula, guqulela ku-mono, shintsha isampula, sebenzisa ama-fade, futhi wenze i-normalize yomsindo.
  • Thumela umsindo njenge-WAV engu-48 kHz, i-OGG, noma i-FLAC ngokucutshungulwa kwangemva.
  • Gcina ulwazi lokukhiqiza kabusha ku-sidecar ye-JSON eduze kwefayela ngalinye lomsindo.
  • Khipha amamodeli ngokuzenzakalelayo ngemva kwesikhathi sokungasebenzi noma uwakhiphe mathupha lapho kudingeka kukhululwe inkumbulo.

Izidingo zesistimu

i-macOS ene-Apple Silicon

  • i-Mac ene-Apple Silicon, njenge-M1 noma entsha.
  • Kunconywa i-RAM engu-16 GB.
  • Cishe u-6 GB wesikhala sediski semodeli ephakathi.
  • git kanye ne-ffmpeg.

i-Windows ene-NVIDIA GPU

  • I-Windows 10 noma i-Windows 11 engu-64-bit.
  • I-NVIDIA GPU ene-VRAM okungenani engu-8 GB kanye ne-driver yamanje.
  • Cishe u-14 GB wesikhala sediski semodeli ne-PyTorch.
  • git kanye ne-ffmpeg.
  • I-akhawunti yamahhala ye-Hugging Face.
< p>Ama-Intel Mac nama-AMD GPU awasekelwa. I-Linux ene-NVIDIA GPU ingase isebenze ne-backend ye-PyTorch, kodwa i-repository ichaza ukusekelwa kwe-Linux njengokungakahlolwa futhi ayinikezi isifaki sayo.

Faka ku-macOS

Okokuqala faka i-ffmpeg uma ingakatholakali:

brew install ffmpeg

Clone i-repository bese usebenzisa isifaki sayo:

git clone https://github.com/kanaworksai-hue/audio-studio-mcp.git
cd audio-studio-mcp
./install.sh

Isifaki singasetha i-uv, sicope i-runtime ye-Stable Audio 3 ehloliwe iye ku-vendor/, sakhe i-virtual environment, bese silanda imodeli ephakathi engu-1.4B. Imodeli icishe ibe ngu-5.9 GB.

Ekugcineni, isifaki siphrinta umyalo wokubhalisa. Sebenzisa umyalo ohambisanayo usebenzisa indlela yangempela eya ku-checkout yakho:

claude mcp add --scope user audio-studio -- /path/to/audio-studio-mcp/.venv/bin/audio-studio-mcp

Faka ku-Windows

Faka i-Git ne-ffmpeg uma kudingeka:

winget install Git.Git
winget install Gyan.FFmpeg

Ngaphambi kokufaka iphrojekthi, vula ikhasi lemodeli ephakathi ye-Stable Audio 3 ku-Hugging Face bese wamukela ilayisensi yayo. Izisindo ze-PyTorch zikhawulelwe, ngakho kufanele unikeze nethokheni yokufinyelela ye-Hugging Face enemvume yokufunda uma isifaki siyicela.

Clone i-repository bese uqalisa isifaki se-PowerShell:

git clone https://github.com/kanaworksai-hue/audio-studio-mcp.git
cd audio-studio-mcp
powershell -ExecutionPolicy Bypass -File install.ps1

Isifaki se-Windows sifaka okuncikile kwe-PyTorch okunikwe amandla yi-CUDA bese silanda i-checkpoint esemthethweni ye-PyTorch. Imodeli ephakathi icishe ibe ngu-10.4 GB. I-Flash Attention ayidingeki; i-runtime ibuyela ku-attention eyakhelwe ngaphakathi ye-PyTorch lapho ingatholakali.

Bhalisa iseva nge-Claude Code usebenzisa indlela ephrintwe isifaki, noma lungisa lesi sibonelo:

claude mcp add --scope user audio-studio -- "C:\path\to\audio-studio-mcp\.venv\Scripts\audio-studio-mcp.exe"

Xhuma amanye amaklayenti e-MCP

I-Claude Desktop, i-Cursor, namaklayenti ahambisanayo angaqalisa iseva kusukela ekucushweni kwawo kwe-MCP. Faka okufakiwe kwe-audio-studio kufayela lokucushwa elifanele, njenge-claude_desktop_config.json noma .cursor/mcp.json:

{
  "mcpServers": {
    "audio-studio": {
      "command": "/path/to/audio-studio-mcp/.venv/bin/audio-studio-mcp"
    }
  }
}

Kuma-Windows, sebenzisa i-executable engaphakathi kwe-.venv\Scripts. Njengoba lokhu kuyi-JSON, ama-backslash kufanele aphindwe kabili:

{
  "mcpServers": {
    "audio-studio": {
      "command": "C:\\path\\to\\audio-studio-mcp\\.venv\\Scripts\\audio-studio-mcp.exe"
    }
  }
}

Yenza isiqeshana sakho sokuqala somsindo

Uma iseva isibhalisiwe, sebenzisa ulimi olujwayelekile kuklayenti lakho le-MCP. Isibonelo:

  • Yenza i-loop ye-lo-fi BGM yemizuzwana engu-60 ku-85 BPM yesigcawu sekhefi, bese unginika abangenele abathathu.
  • Yenza umsindo womnyango osindayo wokhuni oshayeka ephasejini lamatshe, bese uwukhipha njenge-OGG ye-Godot.
  • Yenza umoya omfushane wemvula nokuduma kwezulu wehlathi elingaphandle ebusuku.

Iseva iveza ithuluzi le-generate_audio lokwenza umsindo kusuka embhalweni. Lisekela izinhlobo zomculo, izinsimbi, imiphumela yomsindo, nomsindo ongahluziwe. Emculweni, faka uhlobo, umuzwa, izinsimbi, izendlalelo ezisekelayo, isigqi, nejubane. Emiphumeleni yomsindo, chaza umthombo, isenzo, impahla, nendawo.

Amafayela akhiqiziwe abhalwa ngokuzenzakalela ku-directory ye-outputs/ ye-repository. Ifomethi ejwayelekile ekhiqizwayo yi-WAV ye-44.1 kHz, 24-bit, stereo. Ifayela ngalinye line-sidecar ye-JSON ehambisanayo enemininingwane yokukhiqizwa kwalo.

Imiphumela yomsindo ekhiqizwe yiklayenti le-MCP

Cela abangenele abaningi

Imiphumela ingahluka kakhulu kuye nge-seed. Cela izinguqulo ezimbili kuya kwezine uma ufuna ukuqhathanisa ezinye izindlela:

Yenza abangenele abane bomculo we-adventure we-8-bit onamandla ku-150 BPM, one-lead ekhanyayo ye-square-wave nezigubhu ezinamandla.

Iseva inikeza ama-seed alandelanayo ezinguqulweni. Unganikeza futhi i-seed ethile ukuze uthole umphumela ophindaphindekayo, noma ushiye i-seed bese usebenzisa -1 ukuze uthole i-seed engahleliwe.

Hlela umsindo okhiqiziwe

Sebenzisa i-edit_audio uma ukukhiqizwa kokuqala kuseduze nalokho okufunayo kodwa kudinga ushintsho. Ithuluzi lisekela izindlela ezintathu:

  • Variation: shintsha isitayela saso sonke isiqeshana.
  • Inpaint: khiqiza kabusha ububanzi besikhathi obukhethiwe.
  • Continue: nweba ithrekhi idlule esikhathini sayo samanje.

Isibonelo, ungacela iklayenti ukuthi lelule ithrekhi yemizuzwana engu-20 ibe yimizuzwana engu-40, noma likhiqize kabusha ingxenye enenkinga esikhundleni sokuphinda lidale lonke ifayela:

Qhubeka nale threkhi iye kumasekhondi angu-90, ugcine isitayela namandla afanayo.
Khiqiza kabusha ingxenye esuka kumasekhondi angu-8 kuya kwangu-12 ukuze ushintsho luzwakale lushelela kakhulu.

I-README iphawula ukuthi imiyalelo enesikhathi esiqondile, njengokucela i-climax ngomzuzwana othile, ayilandelwa ngokuthembekile yimodeli. Indlela ethembeke kakhulu ukukhiqiza isiqeshana kuqala bese usebenzisa i-inpaint noma i-continue ukwenza izinguquko zesakhiwo.

Icubungule futhi uyikhiphe imisindo

Ithuluzi le-process_audio lisebenzisa imisebenzi yokuqedela engagudluki. Alikaze liguqule okokufaka kwasekuqaleni. Imisebenzi etholakalayo ihlanganisa ukusika, ukususa ukuthula, ukuguqulela ku-mono, ukuphinda kusampulwe, ukulinganisa umsindo, ukwenza i-loop engenamthungo, ama-fade, ukulinganisa i-peak, nokuguqula okukhiphayo.

Isibonelo:

  • Yenza lo mphumela womsindo ube yi-mono, usike ukuthula, sebenzisa i-fade emfushane, bese uwukhipha njenge-OGG.
  • Yenza i-loop engenamthungo ngalo mculo, sebenzisa i-crossfade ende, bese uwukhipha njenge-WAV.
  • Linganisa le threkhi yomculo ongemuva womdlalo ibe ngu--16 LUFS bese uyikhipha njenge-OGG.

Sebenzisa ukulinganisa kwe-peak emiphumeleni eminingi yomsindo. Ekulinganiseni umsindo, iphrojekthi incoma cishe -14 LUFS ekusakazeni noma kuvidiyo, -16 LUFS emculweni ongemuva womdlalo, kanye -23 LUFS ekusakazeni.

I-peak ceiling ezenzakalelayo ngu--1 dB. Yehlisa ama-peak adlula le ceiling kodwa ayikhuphuli umsindo othule. Amafayela acutshunguliwe agcina imininingwane yokukhiqizwa kwasekuqaleni ngaphansi kwenkambu ethi source kuma-metadata awo.

Gwema ama-click kuma-loop

Sebenzisa i-WAV noma i-OGG uma udinga i-loop engenamthungo. I-MP3 ingangeza i-padding ekuqaleni nasekupheleni, okuvimbela i-loop enembile. Uma kusekhona i-click emhlanganweni, yandisa ubude be-loop crossfade noma cela imodeli ukuthi yenze i-loop ephela emugqeni webha.

Thuthukisa ama-prompt akho

I-Stable Audio 3 isebenza kahle kakhulu ngama-prompt amafushane esiNgisi emisho eyodwa noma emibili, ngokuvamile angeqi cishe amagama angu-45. Gwema amagama ahlobene namazwi afana ne-vocals, singer, noma lyrics, ngoba le phrojekthi ikhiqiza umculo wezinsimbi nemiphumela yomsindo, hhayi ukucula okuqondakalayo.

Isakhiwo se-prompt yomculo

Ukuhleleka okuwusizo kuma-prompt omculo uhlobo nezichasiso, izinsimbi ezihola phambili, izendlalelo ezisekelayo, izigubhu noma isigqi, kanye nomuzwa. Faka i-tempo kupharamitha ye-bpm noma uyifake ngokucacile esicelweni.

Ingoma ye-lo-fi hip-hop chill enepiyano kagesi ethambile, ukuqhekeka kwe-vinyl okuthambile, izigubhu ezine-low-pass filter, ne-bass ethambile edonswayo, enomuzwa okhululekile nowamaphupho.

Isakhiwo se-prompt yomphumela womsindo

Chaza imiphumela yomsindo usebenzisa izici ezine: okudala umsindo, indlela oguquka ngayo, into owenziwe ngayo, nendawo owenzeka kuyo.

umthombo: ibhodlela lengilazi / isenzo: liqhekeka lapho lishayisana, izingcezu zisakazeka / impahla: ingilazi eluhlaza ewugqinsi, umsindo wokuqala obukhali nokhanyayo / indawo: phansi kukakhonkolo, eduze, yomile

Qondanisa ubude besikhathi obuceliwe nokuqukethwe. Imisindo yomthelela ivamise ukusebenza kahle ngemizuzwana eyodwa kuya kwemithathu, izenzo ngemizuzwana emithathu kuya kweyisithupha, umsindo wendawo ngemizuzwana eyisithupha kuya kweyishumi nanhlanu noma ngaphezulu, kanti imijikelezo iba ngemizuzwana eyisithupha kuya kwengamashumi amabili. Amathrekhi aphelele angaba made, imodeli ephakathi isekela imizuzwana efika ku-380.

Amasethingi asemthethweni azenzakalelayo ngu-steps=8 no-cfg=1.0. Uma udinga ukunamathela ngokuqinile kuphrompthi noma ufuna ukusebenzisa iphrompthi ephikisayo, zama inani le-CFG eliphakathi kuka-3 no-7, ukhumbule ukuthi ukukhiqiza kungathatha isikhathi esiphindwe cishe kabili.

Khetha imodeli bese ulungiselela iseva

Imodeli ezenzakalelayo ngu-medium, enikeza ikhwalithi engcono kakhulu futhi esekela ubude besikhathi obufika kumasekhondi angu-380. Amamodeli amancane sm-music kanye sm-sfx ashesha futhi asekele ubude besikhathi obufika kumasekhondi angu-120.

Setha amanani okulungiselela ngomyalo wokubhalisa we-MCP noma ebhulokhini lemvelo leklayenti lakho. Okuguquguqukayo okubalulekile kufaka:

  • AUDIO_STUDIO_OUTPUT_DIR: kushintsha ifolda yokuphumayo. Okuzenzakalelayo ngu-<repo>/outputs.
  • AUDIO_STUDIO_IDLE_UNLOAD_MIN: kukhulula inkumbulo yemodeli ngemva kwenani elishiwo lemizuzu yokungasebenzi. Kusethwe ku-0 ukuze kukhutshazwe ukukhishwa okuzenzakalelayo.
  • AUDIO_STUDIO_SA3_MODEL: kukhetha medium, sm-music, noma sm-sfx.
  • AUDIO_STUDIO_MLX_CACHE_GB: kubeka umkhawulo wenqolobane ye-MLX buffer ku-macOS.
  • AUDIO_STUDIO_DEVICE: kukhetha idivayisi ye-PyTorch efana ne-cuda, cuda:1, noma cpu.
  • AUDIO_STUDIO_BACKEND: kukhetha auto, sa3-mlx, noma sa3-torch.
  • AUDIO_STUDIO_FFMPEG: icacisa ifayela elisebenzisekayo le-ffmpeg uma lingekho ku-PATH.

Yabelana ngenqubo eyodwa yeseva efudumele

Ngokuvamile, iseshini yeklayenti ngayinye iqala inqubo yayo yeseva. Amamodeli alayishwa esicelweni sokuqala futhi akhululwe ngemva kwesikhathi sokungasebenzi. Uma amaseshini amaningi edinga ukwabelana ngenqubo eyodwa efudumele, sebenzisa i-Audio Studio MCP njengeseva ye-HTTP yasendaweni:

.venv/bin/audio-studio-mcp --http 8765 --preload
claude mcp add --scope user --transport http audio-studio http://127.0.0.1:8765/mcp

Ku-Windows, qalisa ifayela elisebenzisekayo elisendaweni ebonakalayo:

.venv\Scripts\audio-studio-mcp.exe --http 8765 --preload

Iseva ye-HTTP ilalela ku-127.0.0.1 kuphela, iyigcine isendaweni kukhompuyutha.

Phatha inkumbulo nokusebenza

Kwi-M1 Mac enenkumbulo engu-16 GB, imodeli ephakathi isebenzisa cishe u-5–7 GB ngenkathi ilayishiwe. Kwi-RTX 3060, yafinyelela cishe ku-5.3 GB ye-VRAM ezilinganisweni zendawo yokugcina ikhodi. Sebenzisa ithuluzi elithi unload_models ukuze ukhulule inkumbulo yemodeli ngokushesha, noma ulungiselele AUDIO_STUDIO_IDLE_UNLOAD_MIN ukuze kuhlanzwe ngokuzenzakalelayo.

Amamodeli amancane anganciphisa kakhulu isikhathi sokukhiqiza iziqeshana ezimfushane. Indawo yokugcina ikhodi ibika ukuthi ukukhiqiza okufudumele kwemizuzwana engu-15 nge-sm-music kuthatha cishe umzuzwana owodwa kwi-M1 Mac futhi cishe imizuzwana engu-0.7 kwi-RTX 3060. Ukusebenza kwangempela kuncike ku-hardware, imodeli, ubude besikhathi, nokuthi imodeli isivele ilayishiwe yini.

Qalisa izivivinyo nokuhlola kokuthuthukisa

Indawo yokugcina ikhodi iqukethe izivivinyo ezingaxhunyiwe ku-inthanethi ezingadingi izisindo zemodeli:

uv run pytest

Ukuze wenze isivivinyo esiphelele ngemodeli yangempela, qalisa isivivinyo sokuhlola. Ku-macOS:

.venv/bin/python scripts/smoke_test.py

Ku-Windows:

.venv\Scripts\python scripts/smoke_test.py

Ikhodi yomthombo ihlelwe yazungeza iseva ye-MCP, ukwakhiwa kwamaphrompthi, ukucutshungulwa kwangemva kokukhiqiza, kanye nokuqaliswa kwama-backend. Inkundla entsha ingangezwa njengenye i-backend ngaphansi kwe-src/audio_studio_mcp/backends/.

Ukulungiselelwa kwephrojekthi ye-Audio Studio MCP noma ukubuka kokuphumayo

Ukucatshangelwa kwelayisensi nokusetshenziswa

Ikhodi ye-Audio Studio MCP inelayisensi ye-MIT. I-runtime ye-Stable Audio 3 ilandwa ngesikhathi sokufakwa, kanti izisindo zemodeli zisatshalaliswa ngaphansi kwe-Stability AI Community License. Buyekeza ilayisensi yemodeli ngaphambi kokusebenzisa umsindo okhiqiziwe ngokwezentengiselwano. I-README ithi, ngesikhathi ibhalwa, ilayisensi ivumela ukusetshenziswa kwezentengiselwano kubantu ngabanye nasezinhlanganweni ezinemali engenayo yonyaka ephelele engaphansi kuka-US$1 million.

Isiphetho

Audio Studio MCP inikeza indlela esebenzayo yasendaweni yokwakha umsindo wemidlalo, izizinda zamavidiyo, umsindo wendawo, nemisindo yomthelela ngezicelo zolimi lwemvelo. Faka i-backend efanele i-hardware yakho, bhalisa iseva neklayenti le-MCP, khiqiza amakhandidethi amaningana, bese usebenzisa amathuluzi okuhlela nokucubungula ukuze uguqule umphumela ongcono ube yimpahla ehlanzekile ye-WAV, OGG, noma FLAC.

Ukuze uthole amasampula nemininingwane yakamuva yokufaka, vakashela indawo yokugcina ikhodi ye-Audio Studio MCP kanye nekhasi lamasampula.