Voxta docs

Music Generation

Services that make songs and instrumentals on your own GPU, for the chat's background music, songs about the story, and the Music Playground.

Music Generation (the MusicGen service type) makes a whole track from a style description and, for a song, tagged lyrics. A track is made, then sent back as a file. Nothing streams while it plays.

Where it is used

  • /music and /song in a chat: background music that fits the scene, or a song about the story, sung and attached to the chat.
  • The Music Playground: write a song, make takes, repaint, remix and mix them.
  • The chat's Gallery, which keeps the chat's tracks to play again or download.
  • Service Tests.

The chat loads the service on its first track and unloads it once the track is done, unless another one is queued. This way the music model does not hold video memory the rest of the chat needs.

Add one

  1. Add Modules → Services, pick a music module and Install, then follow the install steps. All three are experimental and can only be installed by an administrator.
  2. Pick a preset or model, on the Music Generation row of the Configuration page's Music tab. ACE-Step and LeVo install a few presets, by model size. audio.cpp picks its model in the install's Models step.
  3. In a chat, /music and /song use the chat's music generation service, and its summarization model to write the style and lyrics.

Each module's Writing Instructions tell the language model how to write the style and lyrics for that model.

Modules

All run locally and download their models from Hugging Face. ACE-Step and LeVo run in Voxta's own Python, audio.cpp without Python on a CUDA GPU.

ACE-Stepaudio.cpp MusicLeVo 2
LicenseMITMITAcademic, research and education use only
SpeedSeconds per song on a recent GPUSeconds per song on a recent GPUAbout a minute per minute of song
Video memory~5 to ~17 GB, by planner~8 GB, ~13 GB to repaint or remix~12 GB medium, ~22 GB large
Length10–600 s, or let the model choose10–600 sFollows the lyrics
Vocal languages50+50+English, Chinese, Spanish, Japanese
Takes per requestUp to 4, in one passUp to 4, one after the otherUp to 4, one after the other
StemsNoNoVocals and backing
Repaint, extend, remixYesRepaint and remixNo
BPM, key, beats per barYesYesNo
LoRAsYesNoNo

On this page