Text To Song
SongGeneration 2 generates full songs with vocals and instrumentals from tagged lyrics, up to 4 minutes 30 seconds long, styled by caption tags or a reference clip.
Overview
SongGeneration 2, also published as LeVo 2, is a text-to-song model from Tencent AI Lab. You supply lyrics with section tags, and it writes the melody, arrangement and singing for a full song. This page runs the SongGeneration-v2-large checkpoint.
One language model plans the whole song, a second predicts the vocal and accompaniment tracks in parallel, and a diffusion-based music codec turns them into audio. Post-training used supervised fine-tuning and preference optimization (DPO) aimed at musicality and control.
Tencent reports a phoneme error rate of 8.55% for the sung lyrics, plus a listening test in which 20 music professionals rated its songs on six dimensions. It lists Chinese, English, Spanish and Japanese among the lyric languages.
What you can build
- Research comparisons of one lyric sheet rendered under different captions, genre presets or reference clips
- Instrumental pieces for music-generation studies, with the lyrics left empty
- Teaching and research on song structure and lyric alignment
Inputs
- Lyrics (the prompt box): each section tag on its own line with its lines below, and a blank line between sections. [verse], [chorus] and [bridge] carry lyrics. [intro-short], [inst-short], [outro-short] and their -medium forms are instrumental passages and stay empty; Tencent puts short at roughly 0 to 10 seconds and medium at 10 to 20. Closing lines belong in a last [verse] or [chorus], since outros take no words. [pre-chorus] and similar tags are not part of the format. An empty box gives an instrumental piece.
- Caption: comma-separated style tags for gender, timbre, genre, emotion, instruments or BPM, such as "female, pop, energetic, piano". The default is "female, slow, piano".
- Genre: a preset (Pop, Rock, Electronic, Hip-Hop, R&B / Soul, Jazz, Latin, Country, Metal, Ballad, Funk, Soundtrack, World) or Auto.
- Reference audio: one audio file whose style the song follows. Only its first 10 seconds are used, and Tencent suggests a chorus.
A reference clip overrides both genre and caption, and a genre preset overrides the caption.
Outputs
MP3 audio. By default vocals and accompaniment come back as one mixed song. Output Type switches to instrumental only, vocal only, or three separate files: the full mix, the vocal and the accompaniment. The two samples on this page are 48 kHz stereo songs of about 2 minutes 15 seconds and 2 minutes 50 seconds.
Recommended settings
- Temperature defaults to 0.8. Raise it for more variation, lower it for more predictable results.
- Write complete lines and tag every section. Tencent says lyric formatting strongly affects the result.
Limitations
- One song runs at most 4 minutes 30 seconds.
- The model follows a reference clip's timbre and style closely, so the choice of clip changes the result a lot.
- Tencent does not publish a full language list, so test less common languages first.
Safety & compliance
Tencent's license terms for SongGeneration limit use to academic, research and education purposes and rule out commercial or production use. A reference clip can carry a recorded singer's voice, so use recordings you hold the rights to and do not imitate real artists without consent.
Example prompts
Great starting points for SongGeneration 2.
API quick start
Run SongGeneration 2 with a single API call.
{
"prompt": "[verse]\nBright lights flashing in the cit…",
"caption": "female, slow, piano",
"temperature": 0.8,
"inputAudio": "https://your-cdn.com/input.mp3"
}curl -X POST "https://api.wiro.ai/v1/Run/tencent-ailab/songgeneration-2" \
-H "Content-Type: application/json" \
-H "x-api-key: YOUR_WIRO_API_KEY" \
--data-binary @- <<'JSON'
{
"prompt": "[verse]\nBright lights flashing in the cit…",
"caption": "female, slow, piano",
"temperature": 0.8,
"inputAudio": "https://your-cdn.com/input.mp3"
}
JSONOther models you might like
Explore similar music generation models.
