DEV Community

Cover image for Solving the Ground-Truth Bottleneck in Music Source Separation: Announcing mixforge-audio on PyPI
Boobalan Arjunan
Boobalan Arjunan

Posted on Edited on

Solving the Ground-Truth Bottleneck in Music Source Separation: Announcing mixforge-audio on PyPI

🚀 Excited to announce the official release of mixforge-audio (v0.1.0) on PyPI!

If you work in Audio AI, Music Source Separation (Demucs, Spleeter, Open-Unmix), or Bleed Detection, you've hit this core evaluation bottleneck:

👉 On commercial tracks, the true ground truth is completely unknown.

Commercial recordings contain unmeasured mic leakage, mastering compression, and stereo imaging that bias metrics (SDR, SIR, SAR). When models fail, it's hard to tell whether separation failed or if source stems had bleed originally.

mixforge-audio solves this: an open-source Python engine engineered to build reproducible, mathematically exact test mixes with an unalterable answer key and calibrated acoustic bleed.

🔍 What Does mixforge-audio Do?

  1. 🎚️ Exact Superposition: Stems (Vocals, Drums, Bass, Other) mix at chosen or random dB levels. If the mix clips, stems scale down proportionally so the algebraic sum holds identically.
  2. 🎙️ Calibrated Acoustic Bleed: Injects physical mic bleed (e.g., drum bleed into vocals at -26 dB with optional 4th-order Butterworth low-pass at 4 kHz) without corrupting clean reference stems.
  3. 📜 Immutable Answer Key: Writes a cryptographic recipe.json recording every sample offset, gain factor, and clipping ratio.
  4. 🛡️ Math Verification: Automated engine (mixforge verify) verifies |Mixture - ∑ Stems| ≤ ε and checks that leaky stems match recipe bleed.
  5. 🎧 Visual HTML Auditing: Generates a self-contained web player report with multi-track solo/mute capabilities to visually and aurally audit test mixes.

📊 Real-Data Benchmarks (MUSDB18-HQ)

Validated across 20 multi-track songs (40 mixes) from MUSDB18-HQ:

✅ Verification Pass Rate: 40 / 40 mixes OK (0 failures)
✅ 64-Bit Float Precision: Max |Mixture - ∑ Stems| = 6.52 × 10⁻⁸
✅ Flat Bleed Precision: 0.0000 dB gap between requested and measured leakage
✅ Low-Pass Bleed: -14.45 dB high-frequency damping above 4 kHz
✅ Official Mix Residual: -69.3 dB median vs original master mixes (well below the human auditory threshold)

⚡ Zero Heavy ML Bloat

No PyTorch, TensorFlow, or GPU compute needed. Runs purely on NumPy, SciPy, and SoundFile for instant, sub-second execution on any CPU or CI/CD runner.

📦 Quick Start & Links

Install directly from PyPI:
pip install mixforge-audio

Generate a test dataset with bleed in 1 command:
mixforge make stems/ --out mixes/ --count 3 --seconds 10 \
--bleed "drums>vocals:-26@4000" --bleed "bass>other:-30" \
--seed 42

Verify mix integrity:
mixforge verify mixes/ --html report.html

🔗 PyPI: https://pypi.org/project/mixforge-audio/
💻 GitHub: https://github.com/BoobalanArjunan/Mix-Forge

If you're benchmarking separation models or training audio bleed detectors, give it a spin! ⭐

AudioML #MachineLearning #MusicAI #DeepLearning #SourceSeparation #AudioEngineering #Python #OpenSource #SignalProcessing #MIR #ArtificialIntelligence


Enter fullscreen mode Exit fullscreen mode

Top comments (0)