Per-stem melody + stems dB levels + meter (levels 0.72 -> 0.79 on the new scorer) - #8
Open
DheerajNalapat wants to merge 2 commits into
Open
DheerajNalapat wants to merge 2 commits into
DheerajNalapat wants to merge 2 commits into
Conversation
The new scorer's ev_meter reads grid.beats_per_bar; we didn't claim it ('no
beats-per-bar claimed'). Emit it from the time signature (4 for 4/4). Necessary
but not sufficient: meter also needs chord changes on the grid, which the chords
work will supply.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…ion enabler New StemSeparationAnalyzer runs demucs once (cached to work/stems_cache/ by source) and publishes the waveforms in ctx. MelodyAnalyzer now pitch-tracks the isolated vocals stem into note events (mix contour was 0.15; vocals -> 0.82). StemsAnalyzer now emits per-bar dB levels vs the mix RMS (comparable), which sum to the mix as the scorer requires (normalised-per-stem was 0.66 -> dB 0.94). port passes the comparable stems through and marks guitar/piano absent (not 0 dB). Experiments on levels: melody 0.15->0.82, stems 0.66->0.94. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Continues closing the gap to the
listen-produced model maps under the updated scorer (which addedstems/meter/tempo_stability). Model targets are now levels 0.90 / the-nights 0.76.What this adds
work/stems_cache/(keyed by source), and publishes them inctx. The enabler for every per-stem field. Cached runs skip demucs (0.02s).[start,dur,midi,name]. The mix contour was 0.15 (the dominant mix pitch isn't the vocal); the stem gives 0.82.comparable:true) — summed as energy they track the mix, which the newev_stemschecks. Normalised-per-stem was 0.66; dB → 0.93.ev_meter(prerequisite; meter also needs chord segments).Result — levels (deep, current scorer)
0.72 → 0.79: melody 0.15→0.82, stems 0.66→0.93, plus grid 0.88 / bars 1.00 / sections 1.00 / energy 0.86 / pump 0.96 / tempo_stability 0.83. 56 tests pass.The remaining gap to 0.90 — gated on infra not installed here
attack.pyattack-precision (the model tags 8 instruments with ~3520 precisely-placed hits). Our 4-stem mix onsets score worse.tools/setup.shclones but isn't present in this environment.So melody + stems are the wins reachable with the current stack; the last ~0.11 needs the ChordMini + 6-stem machinery.
🤖 Generated with Claude Code