- Home
- AI Singing Voice Generators
- SoulX-Singer
SoulX-Singer
What it is
SoulX-Singer is an open-source model for synthesizing singing voices for singers it has not been tuned on. It supports melody-conditioned F0 contours and score-conditioned MIDI notes to control pitch, rhythm, and expression. Its voice-conversion system can transform raw singing audio into a target singer’s voice while retaining melody, rhythm, and lyrics, without requiring lyric or MIDI transcriptions. The system supports Mandarin Chinese, English, and Cantonese, along with singer-timbre cloning, cross-lingual synthesis, and lyric editing while preserving natural prosody. Preprocessing tools handle vocal separation, dereverberation, F0 extraction, voice activity detection, and lyric and note transcription. A MIDI editor can adjust lyrics, phoneme alignment, pitches, and durations before synthesis. Soul-AILab provides singing-generation, vocal-conversion, and MIDI editor demos through Hugging Face Spaces; the repository also supports local inference using Conda, Python 3.10, pip-installed dependencies, and web UI scripts. The code and model weights use the Apache 2.0 license. Maintainers caution that automated preprocessing may misalign singing with lyrics and notes, so manual correction may be needed.
For and against
For
- Works with singers without per-speaker fine-tuning.
- Offers F0 and MIDI-based control.
- Supports Mandarin Chinese, English, and Cantonese.
- Includes tools for vocal and note preprocessing.
- Code and weights use the Apache 2.0 license.
Against
- Automatic preprocessing may misalign audio with lyrics and notes.
- Local inference requires Conda, Python 3.10, and dependencies.
- Maintainers prohibit unauthorized impersonation or deceptive audio.
The verdict
SoulX-Singer offers synthesis, conversion, and MIDI editing tools under an open-source license. Users should account for possible preprocessing misalignment and follow the project’s consent and impersonation restrictions.
Who it suits
SoulX-Singer is for researchers and developers working with singing voice synthesis or conversion. Its online demos offer an access route, while local inference is available through the repository setup.
Plans and prices
SoulX-Singer
Researchers and developers are free to use the code and model weights · Apache 2.0 license
Read 1 Oct 2026 · github.com
On the record
- Control modes
- It supports melody-conditioned F0 contour control and score-conditioned MIDI note control for pitch, rhythm, and expression.github.com · read 1 Oct 2026
- Core function
- SoulX-Singer is a high-fidelity zero-shot singing voice synthesis model for generating realistic voices for unseen singers.github.com · read 1 Oct 2026
- Dataset scale
- The stated training dataset contains more than 42,000 hours of aligned vocals, lyrics, and notes.github.com · read 1 Oct 2026
- Deployment
- The repository can be cloned, installed with Conda and pip, and run locally through Python web UI scripts.github.com · read 1 Oct 2026
- Editing and cloning
- Features include singer-timbre cloning, cross-lingual synthesis, and lyric editing while preserving natural prosody.github.com · read 1 Oct 2026
- Languages
- The system supports Mandarin Chinese, English, and Cantonese.github.com · read 1 Oct 2026
- License
- The code and model weights are released under the Apache 2.0 license for researchers and developers to use.github.com · read 1 Oct 2026
- License and safety
- The project uses Apache 2.0 and asks users to respect intellectual property, privacy, and consent and avoid unauthorized impersonation or deceptive audio.github.com · read 1 Oct 2026
- Local deployment
- The repository supports local inference through Conda with Python 3.10, pip-installed dependencies, and local WebUI scripts.github.com · read 1 Oct 2026
- MIDI editing
- A MIDI Editor supports editing lyrics, phoneme alignment, note pitches, and durations before inference.github.com · read 1 Oct 2026
- MIDI integration
- Generated metadata can be exported to MIDI, edited for lyrics, phoneme alignment, pitches, and durations, and imported back for synthesis.github.com · read 1 Oct 2026
- Model distribution
- Pretrained synthesis, conversion, and preprocessing models are downloaded through Hugging Face Hub commands.github.com · read 1 Oct 2026
- Notable limitation
- The maintainers warn that automatic preprocessing may misalign singing audio with lyrics and notes and recommend manual correction.github.com · read 1 Oct 2026
- Online access
- Soul-AILab provides a running SoulX-Singer demo on Hugging Face Spaces and a separate running MIDI Editor Space.huggingface.co · read 1 Oct 2026
- Pitch and score control
- It supports melody-conditioned F0-contour control and score-conditioned MIDI-note control for pitch, rhythm, and expression.github.com · read 1 Oct 2026
- Preprocessing
- Its preprocessing toolkit performs vocal separation and dereverberation, F0 extraction, voice activity detection, lyrics transcription, and note transcription.github.com · read 1 Oct 2026
- Purpose
- SoulX-Singer is a high-fidelity zero-shot singing voice synthesis model for generating realistic voices for unseen singers.github.com · read 1 Oct 2026
- Support
- The project lists three contact emails and invites technical discussion through WeChat or Soul app groups.github.com · read 1 Oct 2026
