Skip to content

StemgenRT 0.7.0

Latest
Latest

Choose a tag to compare

View all tags
@axeldelafosse axeldelafosse released this 11 Oct 21:44
v0.7.0
85e715f
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

StemgenRT 0.7.0 replaces the joint separation model with independently trained Drums, Bass and Vocals models in one FP32 ONNX graph. Other is the aligned input minus those three estimates.

The plug-in remains named StemgenRT, retaining its existing host IDs, controls and output routing. Updating replaces the model used by existing StemgenRT instances. The temporary StemgenRT Independent test build used separate IDs; replace those instances explicitly if needed.

Changes

  • Three final 40,000-update EMA models, with thirteen persistent states and a self-contained 32 MB graph.
  • Cleaner vocals, better silence behavior and less vocal bleed in drums and bass in the maintainer's listening comparison. Other can contain stronger residual errors.
  • Stereo 44.1 kHz; 256 samples / 5.805 ms of graph-plus-worker latency at a 128-sample host buffer. Larger buffers include their additional reported reserve.
  • Complete matching model and training source, also bundled in the plug-in source archive. Draft papers and private experiment records are excluded.

Validation

Independent PyTorch waveform/state parity and streaming lifecycle checks passed. Matching model/training code passed 301 package tests, plus the retained frozen-source and integration suites. The final release packages were built from 85e715f1a1982b29204e11595d6fa3f0bde9810a in GitHub Actions run 38175830082. macOS ARM64 and Windows x86-64 builds and test suites passed, including strict Mac bundle signature/linkage checks. Archive checks verified version 0.7.0, the established StemgenRT identity, native architectures, embedded runtime and the exact model in every package. The full deployment evaluation also reproduced the native models’ quality scores.

Physical M4/M4 Pro sustained timing and DAW deadline measurements remain unmeasured for these weights. The latency figure describes algorithmic delay, not CPU deadline qualification.

Downloads and installation

  • macOS AU/VST3: Apple Silicon, macOS 14+. The complete bundles belong in ~/Library/Audio/Plug-Ins/Components/ and ~/Library/Audio/Plug-Ins/VST3/, respectively.
  • Windows VST3: x86-64; use your standard VST3 directory.
  • Source: the source archive includes the real ONNX file and matching training code; a Git checkout requires Git LFS. Build/install instructions are in README.md.

Close the DAW before replacing the complete old bundle, then restart it. Keep the previous version outside the plug-in directory if needed for rollback. The model and ONNX Runtime 1.26.0 are included. macOS bundles are ad-hoc signed, without Developer ID signing or notarization. Verify downloads with SHA256SUMS.txt.

Assets 8
Loading