- Return an unchanged copy for a non-finite/zero grid (NaN downbeat used to
delete every guitar note).
- Skip an added note when a same-pitch note (bar, neighbouring bars or already
added) lies within tol; avoids near-duplicates across bar lines.
- bench run/real: --no-repeat-vote passes { repeatVote: false } to the pipeline.
- Tests: notes before a late downbeat, bpm 0 / NaN / Infinity, trailing partial
bar, half-tempo grid, neighbour-bar duplicate, pipeline-level on/off.
Eval unchanged: synthetic t1a 0.885 t1b 0.840 t1c 0.752 t1d 0.737, lnoe 0.507.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Bars whose transcriptions are similar (16th-slot occupancy cosine >= 0.6,
diagonal-smoothed, and exact pitch/onset match >= 0.3) are treated as repeats.
Notes recurring in < 15% of a bar's repeats are dropped; notes in >= 25% and
>= 3 repeats but missing from the bar are added. Default on for guitar.
Synthetic median guitar F1 t1a 0.866->0.885, t1b 0.814->0.840,
t1c 0.734->0.752, t1d 0.721->0.737; lnoe 0.403->0.507.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Demucs-6s scatters guitar energy across its guitar, piano and other stems.
Synthetic t1 median guitar F1 rises from 0.39-0.60 to 0.72-0.87; the real
'Last Night on Earth' reference goes 0.390 -> 0.403.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Bass absolute SDR floor, guitar/piano order pinned by stem energy,
finite + non-silent checks on the mono and 3 s slice paths, and the
test.failing comment cites ruling R7.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Radix-2 FFT and torch-compatible stft/istft (hann, center, reflect,
normalized). HTDemucs-6s ONNX separator mirroring demucs apply_model:
whole-mix normalisation, 343980-sample centred segments, 25% overlap
with triangular weights, _spec/_ispec around the graph, spectral +
time branch sum. Brief's guitar/bass > 3 dB criterion is unmet by the
model on t1b/1 and is kept as test.failing pending a ruling.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
