# Listening Feedback Backlog This is the working backlog for ear-led Music Stack improvements. Use `listen.html` for the actual listening pass, then turn short feedback into small implementation tasks. ## Operating Rule - Keep `listen.html` separate from dashboards. - `listen.html` is for listening, cues, and unresolved sound notes. - OpenClaw / review queue is for tracking decisions and next actions. - Short feedback is enough if it names the target and the audible problem. Feedback shape: ```text target: FM funk / 60 sec heard: drums improved, melody still repeats keep: clavi color, pocket fix: reply phrases, rests, wider register priority: high ``` ## Current Reopened Notes ### 2026-06-01 phone notes - `FM funk`: sounded packed/crushed. v299 pass should be checked for more headroom, less sub buildup, and less tape/limiter flattening. - `Band Room AI`: playback can solidify over time. v299 pass throttles UI telemetry and adds foreground AI scheduler stall recovery; needs a real phone/PWA listen to confirm. ### Hazama FM - Melody can still read as a short repeated fragment, even after the melodic director work. - Rhythm and mode/key changes previously felt awkward or badly timed; phrase gating exists, but the ear check should stay open. - Bass can still fall into root / fifth / octave loop behavior. - `lofi`, `jazz`, and `funk` pocket timing has been improved; the next question is whether the music develops enough to stay interesting. - `techno` should expose acid motion without bright EDM fatigue, constant hats, or limiter-crushed low end. - `piano` should be a foreground object, not hidden metadata under the engine bed. - `ambient` should remain safe air and not become generic dark pad wash. ### Band Room - AI recreation can still feel thin or uneven compared with original stems. - AI mode can read as "mostly drums" if bass / voice / chords do not carry enough body. - Voice / melody can feel stuck or too synthetic if it shadows the same contour too often. - Practice use should prioritize stable click / drums / bass before decorative polish. - Car / Bluetooth volume and route behavior still needs real-device listening validation. ### Music Core Rig - AUTO MIX should not keep increasing density until the piece loses space. - Low-end pressure, bright transient build-up, and reference-like detail should be judged by ear, not only by runtime metadata. - Reference spread should be audible as Music-specific behavior, not just docs or console state. ## Worker-Gaming Run Log ### 2026-06-01 - tabasco/human-fly v301 recreation-cycle Command: `C:\workspace\music-stack-worker\.venv\Scripts\python.exe -X utf8 scripts\worker-gaming-pipeline.py --worker-root C:\workspace\music-stack-worker recreation-cycle tabasco human-fly --with-analysis --with-drum-candidate --force` Artifacts stay outside Git until human listening approves them: - Report: `C:\workspace\music-stack-worker\reports\tabasco\human-fly\recreation-cycle-20260601-132113\recreation-cycle-report.md` - AI recreation stems/mix: `C:\workspace\music-stack-worker\ai-recreation\tabasco\human-fly\` - Drum candidate: `C:\workspace\music-stack-worker\reports\tabasco\human-fly\recreation-cycle-20260601-132113\drum-frames-tabasco-human-fly.candidate.json` Measurement result: - `pass_basic_audio_check: true`. - `mix.wav`: duration `239.491s`, peak `0.85065`, centroid `3044.9 Hz`, DR `11.79 dB`. - This clears the v301 numeric gates: centroid moved down from `3785.0 Hz` to under `3200 Hz`, and DR moved from `16.49 dB` into the `8-13 dB` target window. - Bass is no longer buried against drums (`bass` RMS `0.05703` vs `drums` `0.05595`). `other` is still quieter (`0.04197`) but now has a stronger body bed for A/B listening. Next ear check: - Human Fly AI should sound less bright and less like drums-only. - Confirm the new glue does not make intro/break sections feel flat. - Confirm `other` is present as guitar/chord body, not only pad wash. - Keep the drum-frame candidate review-only until the groove is approved by ear. ### 2026-06-01 - tabasco/human-fly recreation-cycle Command: `C:\workspace\music-stack-worker\.venv\Scripts\python.exe -X utf8 scripts\worker-gaming-pipeline.py recreation-cycle tabasco human-fly --with-analysis --with-drum-candidate` Artifacts stay outside Git until human listening approves them: - Report: `C:\workspace\music-stack-worker\reports\tabasco\human-fly\recreation-cycle-20260601-111702\recreation-cycle-report.md` - AI recreation stems/mix: `C:\workspace\music-stack-worker\ai-recreation\tabasco\human-fly\` - Drum candidate: `C:\workspace\music-stack-worker\reports\tabasco\human-fly\recreation-cycle-20260601-111702\drum-frames-tabasco-human-fly.candidate.json` Findings to solve in the next Band Room pass: - AI mix passed basic audio generation, but it is brighter than target (`3785 Hz` centroid vs `2402 Hz` target). - AI mix has much wider dynamics than the source (`16.49 dB` vs `8.73 dB`), so it can read as less glued even before style judgement. - Bass and other stems are quiet against drums (`-29.1 dB` and `-34.0 dB` RMS vs drums `-23.4 dB`), matching the "thin / mostly drums" note. - Drum-frame candidate has source-derived events for 8 sections; promote it only after listening proves the extracted ghosts/crashes serve practice use. - Ableton on worker-gaming still needs VST3 rescan for Kontakt. Sonar is the DAW path for this pass; EP-133 is visible, UR44 capture belongs on the Intel studio PC with Yamaha Steinberg ASIO. ## Implementation Translation - "Melody is boring" -> add phrase-level call/answer, rests, register changes, and stronger contour variation before changing timbre. - "Pocket feels wrong" -> inspect drum-frame microMs, governor amounts, and tempo handoff before changing volume. - "AI recreation is not together" -> first balance drums / bass / voice / chords, then check section-aware arrangement. - "Too loud or harsh" -> check output, bus gain, compression/limiter pressure, and high-frequency sources. - "Good part exists" -> preserve that source or role before making broad sound changes.