Voxta docs

LeVo 2

Songs with vocals made on your own GPU by Tencent's LeVo 2 (SongGeneration 2), from lyrics and style tags.

LeVo 2 (Tencent's SongGeneration 2) makes whole songs with vocals from lyrics and a few style tags. It can give the vocals and the backing as separate stems.

Its license allows academic, research and education use only, not commercial or production use. Experimental and admin-only.

Setup

Add the service

Add Modules → Services → LeVo 2 → Install, then follow the install steps. Voxta extracts LeVo's inference code into the models folder and downloads the shared runtime (~6 GB). The song model is downloaded the first time a preset that uses it loads. Watch the Terminal for progress.

Pick a preset

PresetSong modelDownloadVideo memory
Mediumv2 medium~7 GB~12 GB
Largev2 large~13 GB~22 GB

Large sings and arranges better.

Settings

SettingNotes
Models Directory (module)LeVo's folder: its code, the runtime and the song models. Default Music/LeVo under your models folder.
Song Modelv2 medium or v2 large.
Style Writing InstructionsHow the language model describes the music for LeVo: a few short comma-separated tags (singer, genre, mood, main instruments), never a sentence.

What it takes

LengthFollows the lyrics, up to 4 min 30 s. It cannot be set.
StyleTags, up to 300 characters.
LyricsUp to 2400 characters.
Vocal languageEnglish, Chinese, Spanish or Japanese, read from the lyrics.
TakesUp to 4, made one after the other.
StemsVocals and backing as two tracks.

An instrumental is the song's structure without words, rendered as the backing track. LeVo has no BPM, key, repaint or remix.

It is slow: about a minute per minute of song on a fast GPU. Stems render the song's audio a second time, with the other part silenced, which adds some time per take.

On this page