StemForge
github.com
- Free plan
- Open source
- Windows
- macOS
- Linux
- Web
- Self-hosted
What StemForge is
StemForge is an open-source audio workstation that runs locally in a browser and connects workflows for stem separation, MIDI extraction, audio generation, composition, mixing, and export. Demucs and BS-Roformer separate audio into vocal, drum, bass, other, guitar, and piano stems; spectral analysis helps recommend an engine and model. MIDI extraction covers instrument stems and vocal melodies, and Web MIDI can send MIDI to hardware synths. The Synth tab uses Stable Audio Open for text-conditioned generation up to 600 seconds, with optional audio or MIDI conditioning. Compose uses AceStep for song generation and RVC for voice conversion; Mix brings stems, MIDI-rendered tracks, synth outputs, and composed songs together. Exports include WAV, FLAC, MP3, and OGG. The project says audio stays on the user's computer and internet access is unnecessary after initial model downloads. Setup requires Python 3.12, FFmpeg, and FluidSynth with a GM soundfont. NVIDIA acceleration requires driver 580 or newer; CPU-only operation is significantly slower. Models download on first use, and AceStep models require about 20 GB.
Who it suits
StemForge suits audio creators who want locally run tools for separating stems, extracting MIDI, generating or composing audio, and mixing outputs. Users need the listed setup dependencies and suitable hardware for GPU acceleration.
Good to know, and what to watch
- Connects separation, MIDI, generation, composition, and mixing workflows.
- Exports WAV, FLAC, MP3, and OGG.
- Audio remains local, according to the project.
- Supports Web MIDI connections to hardware synths.
- No subscriptions or per-track fees are listed.
- Setup requires Python 3.12, FFmpeg, and FluidSynth.
- CPU-only operation is significantly slower.
- AceStep models require about 20 GB.
- Stable Audio Open requires license acceptance and authentication.
Plans
Every plan and priceLocal StemForge
StemForge
Open source
The verdict
StemForge brings multiple audio workflows together while keeping audio local, according to the project. Before setup, account for its dependencies, model downloads, and the GPU requirement for acceleration.
What the maker publishes
- AMD support
- Linux AMD GPUs use a separate ROCm 7.1 project variant, but the project describes end-to-end AMD hardware support as untested. source
- Audio generation
- The Synth tab uses Stable Audio Open for text-conditioned audio generation up to 600 seconds, with optional audio or MIDI conditioning. source
- Audio generation
- Stable Audio Open is used for text-conditioned stereo generation with optional stem, MIDI, and mix conditioning and project support up to 600 seconds through chunking. source
- Basic Pitch integration
- Spotify describes Basic Pitch as an audio-to-MIDI converter with pitch-bend detection and publishes it under Apache-2.0. source
- Documentation
- The repository links an INSTRUCTIONS.md guide, requirements, installation steps, launcher flags, model-cache guidance, and a third-party notices file. source
- Export formats
- Export supports WAV, FLAC, MP3, and OGG, including ZIP downloads. source
- Hugging Face gate
- The Synth tab requires accepting Stable Audio Open's license and logging in locally with a read token before model access. source
- Install
- Installation uses a recursive Git clone, uv sync, and uv run python run.py, then serves the interface at localhost port 8765. source
- Integrations
- The documented workflow connects Separate, MIDI, Synth, Compose, Mix, and Export tabs so outputs can flow between pipelines; MIDI Out connects to hardware synths via the browser's Web MIDI API. source
- Licence
- The repository identifies StemForge itself as Apache License 2.0 and permits commercial use, modification, and distribution. source
- Limit: Demucs artifacts
- Demucs documentation reports substantial bleeding and artifacts for the piano source in its experimental six-source model. source
- Limit: Demucs maintenance
- The Demucs repository is archived and says its original maintainer no longer actively maintains it beyond important fixes. source
- Limit: licence inconsistency
- The v1.0.0 release page instead describes that release as personal and non-commercial under PolyForm Noncommercial 1.0.0, creating a material licence discrepancy. source
- Limit: Stable Audio language and style
- The model card says performance is weaker outside English, varies across musical styles and cultures, and is generally better for sound effects and field recordings than music. source
- Limit: Stable Audio vocals
- Stability's model card says Stable Audio Open is not able to generate realistic vocals. source
- Limit: third-party rights
- The project warns that its Apache licence does not cover third-party model weights and lists some weights as non-commercial, unknown-licence, or requiring separate review. source
- Local processing
- The project says audio remains on the user's computer and that internet access is unnecessary after model downloads. source
- macOS support
- Apple Silicon Macs use MPS acceleration, while Intel Macs operate CPU-only. source
- Maker identity
- The repository owner identifies as Todd (tsondo) and lists Germany as the location. source
- MIDI
- StemForge extracts MIDI from instrument stems with BasicPitch and vocal melodies with faster-whisper and pitch tracking, and can send MIDI to hardware synths through Web MIDI. source
- MIDI browser support
- MIDI Out requires Chrome or Edge and a secure browser context; the project documents HTTPS or localhost as supported secure origins. source
- MIDI extraction
- Instrument MIDI uses Basic Pitch, while vocal MIDI uses faster-whisper with pitch tracking; FluidSynth provides browser previews. source
- MIDI Out
- Extracted MIDI can be sent to software or hardware synthesizers through Web MIDI, with program changes and a panic control; Chrome or Edge is required by the project. source
- Mixing
- Mix combines audio stems, MIDI renders, synth outputs, and composed songs with per-track instruments and volume, then renders FLAC. source
Checked against the maker's pages on 15 Oct 2026.