Main FlowStrument representation-floor page

<aside>

This gate removes the neural codec from the audio path. Each embedded panel compares the source prompt, a real same-instrument target note, direct measured Rubber Band R3, PC-NSF representation transfer, and the static HyperGANStrument baseline. Use the white Play buttons; the first click may briefly show Loading.

</aside>

Result

Direct R3 has the best pooled real-target timbre proxy across nine local rows: 4.28 dB LSD, versus 4.95 dB for PC-NSF transfer and 5.53 dB for HyperGANStrument. This is a listening prerequisite, not a promotion: the real target is a different performance, and the pooled pitch referee remains tracker-sensitive.

Decision

Do not train another neural decoder or waveform post-filter. Keep the transported waveform intact. The next universal flow should learn pitch-dependent smooth timbre-envelope transport from real same-instrument notes, while R3 retains phase, transients, decay, and the measured source contour.

Keyboard

Guitar

Other acoustic coverage

Notes

• HyperGANStrument remains evaluation-only; the local bundle license is UNVERIFIED.
• Two additional local rows are available in the repository but could not be attached because the Notion connector reached its workspace usage limit.
• Local result: docs/direct_r3_acoustic_transport_gate.json.