Full-song key estimation test; one global label erased section-level modulation; this project estimates likely key regions instead of forcing one global key label on a whole song

Record

Problem

One global key label loses section-level modulation

Bottleneck

Keep region output visible while moving a large checkpoint out of git history

Fix

Two-stream chroma/HPCP inference, CLI/API paths, release-asset checkpoint

Result

Runnable inference surface with SHA-256 verified checkpoint loading

Guardrail

Benchmark page waits on training manifest and dataset list

Built

  • two-stream Transformer model for chroma and HPCP-style harmonic features
  • local WAV and YouTube URL inference paths
  • CLI and FastAPI server
  • release-asset checkpoint with SHA-256 verification
  • manifest-based evaluation script

Signals

  • output exposes region start/end, key name, confidence, and probabilities
  • checkpoint is distributed as release asset, not repo history
  • local manifest evaluator exists for exact-key accuracy
  • training provenance remains the next missing record
region output

Keeps section boundaries visible instead of hiding changes behind one label

confidence

Model certainty per detected key region

checkpoint

Large model file is loaded from release assets with SHA-256 verification

missing record

Training data and training run manifest still need cleanup before stronger claims