Releases: tsondo/StemForge
StemForge v1.0.0
StemForge is a hybrid generative workstation and stem-driven remixing application for producers, sound designers, and AI music researchers. It combines classic source separation with modern generative pipelines in a single desktop workflow — no cloud required, no subscriptions.
What it does
Separate · Demucs HTDemucs / MDX neural source separation. Splits any audio into vocals, drums, bass, and other stems at 44.1 kHz.
MIDI · BasicPitch polyphonic transcription for instrument stems; faster-whisper + PYIN pitch tracking for vocal stems. Onset sensitivity and sustain threshold are tunable per run. Output is a merged multi-track MIDI file plus per-stem players.
Mix · Multi-track mixer with per-track volume (dB), mute, solo, GM instrument assignment for MIDI tracks, and a master render to FLAC. Supports audio stems, MIDI tracks, and generated audio side-by-side.
Generate · Stable Audio Open diffusion model for text-to-audio synthesis. Vocal Preservation (VP) Mode conditions generation on a live vocal stem with tunable conditioning strength, timing lock, and negative prompt — useful for style transfer, re-harmonisation, and guided generation without fine-tuning.
Export · Transcode any combination of pipeline outputs (stems, MIDI, generated audio, mix render) to WAV, FLAC, MP3, or OGG in a single pass.
Technical notes
- Runs fully offline after first model download; no API keys required for core features
- GPU acceleration via PyTorch CUDA (optional); falls back to CPU
- DearPyGUI desktop UI with animated waveform and MIDI visualisers, click-to-seek, per-visualiser colour-coded hover themes
- Thread-safe pipeline execution; UI remains responsive during inference
- Ace-Step JSON sidecar support — BPM, key, time signature, and caption auto-filled from compatible generation metadata
License
Free for personal and non-commercial use under the PolyForm Noncommercial License 1.0.0.
Commercial licensing available — contact tsondo@gmail.com.
Copyright © 2026 Todd Green