SoulX-Singer
Soul AI Lab · Released Feb 2026
SoulX-Singer is Soul AI Lab's open-source, Apache 2.0 singing model: zero-shot singing voice synthesis from a MIDI score or melody, plus singing voice conversion (SVC, added in March 2026), all from a short voice prompt with no per-voice training. In our runs on a 6 GB RTX 3060 Laptop GPU, score mode sang every line of a written jingle right within four takes. Conditional because takes are not deterministic, SVC blurs hard consonants, and SVC on a text-to-speech vocal makes the words worse.
Is it right for you?
Good for
- Singing a written melody in a cloned voice: score mode takes MIDI notes and lyrics and sang every line right within four takes in our runs
- Zero-shot voice cloning from a voice prompt of about 10 seconds, with no per-voice training run
- Re-singing an existing vocal in another voice with SVC while keeping its timing and melody
- Consumer hardware: score mode and SVC both ran on a 6 GB RTX 3060 Laptop GPU, one job at a time
- Commercial projects that need a permissive license: the code and weights are Apache 2.0
Not good for
- Reproducible output: the model samples, so the same input sings differently on every run; keep the audio of every take you pick
- Converting a text-to-speech vocal with SVC: in our runs it made the words worse on every prompt
- Lyrics with dense hard consonants through SVC, which keeps the flow but blurs consonants
- Putting new lyrics on a borrowed pitch curve in melody mode, which came out unintelligible in our runs; use score mode
Pricing
Input
Free (self-hosted open weights)
Output
Free (self-hosted open weights)
Context
Runs locally, 6 GB GPU tested; Apache 2.0
No verdict changes yet
The clock starts day one. Changes land here as our verdict evolves.
Sources
Verification log
No verification checks yet
We haven't logged a verification check for this entry. Once a check runs, its history shows here.