Twelve samples comparing Piano2Pop (PiCo) with MuseControlLite and ACE-Step, all KaraoKeysPH covers from the 476-song evaluation panel. Each sample links its 30-second window in the piano cover the models were given and in the original song, both on YouTube: the recordings are copyrighted, so only model outputs are hosted here. The original song includes vocals; the models generate the vocal-free accompaniment.
Every input uses the same fixed text prompt, “instrumental, no vocals”; the piano is the only song-specific condition. Model outputs are matched to −16 LUFS integrated with a −1 dBTP true-peak ceiling. Please listen with headphones.
What we hear. PiCo, trained on pair-supervised piano covers, usually sounds the fullest and most transformed, with the richest low end, in line with its best real-piano APA, FAD and spectral-shape scores in the paper; it can, however, sound less clear, at times muddy. Among the pair-free systems, FST is the most balanced: it more often brings in drums, a clear bass line and effects. Base and RDA more often stay close to the piano and lack low-frequency content and drums, consistent with the paper’s finding that both under-produce bass and drums, and RDA brings no consistent gain over Base, occasionally sounding worse. The differences among the four systems are not always significant, but they are audible.
Base, FST, RDA, and PiCo on eighteen samples, with the authors’ listening comments where we have them. Each links the piano cover the models were given at the window’s start; every input uses the fixed text prompt “instrumental, no vocals”. Model outputs are matched to −16 LUFS integrated with a −1 dBTP true-peak ceiling.
How the four systems choose their training transcriptions and apply augmentation, followed by evaluation results and paired comparisons. Training choices are made separately for each four-bar segment.
Ratios are nominal draw probabilities. If a selected transcription is missing, too sparse, or exceeds 25 simultaneous notes, the loader tries other eligible variants and ultimately falls back to the baseline, so realized ratios can differ slightly.
We first separate each training instrumental with HTDemucs, then transcribe one of four function-focused stem mixtures. The mixtures are weighted equally within the function-selective branch:
| Stem mixture | Included stems | Within branch | All FST draws |
|---|---|---|---|
| Full | Drums, bass, other, guitar, piano | 0.25 | 0.55 |
| Harmonic | Bass, other, guitar, piano | 0.25 | 0.15 |
| No-bass | Drums, other, guitar, piano | 0.25 | 0.15 |
| Piano + bass | Piano, bass | 0.25 | 0.15 |
In FST, each four-bar segment draws the original full-instrumental piano transcription with probability 0.40 or the function-selective branch with total probability 0.60. Each mixture contributes 0.60 × 0.25 = 0.15 through the function-selective branch; because the normal transcription is also Full, its aggregate ratio is 0.40 + 0.15 = 0.55. The chosen MIDI is used without further style modification.
RDA always starts from the original baseline transcription. As reported in the paper, it uses two augmentation types. Each four-bar segment is left unchanged with probability 0.50; otherwise it independently draws one of the following:
| Operation | Draw ratio | Exact transformation |
|---|---|---|
| Function-aware note deletion | 0.20 | Keep each onset’s bass note. Other chord-function notes draw a 0.05–0.20 deletion rate; tension/passing-function notes draw 0.20–0.60. |
| Octave folding | 0.30 | Shifts non-bass notes one octave toward the local register while preserving the bassline and topline boundaries: an eligible note moves ±12 semitones, at a per-onset rate drawn from 0.10–0.50, only if it stays between the bass and the top voice. |
These are the two augmentation families named in the paper. The function labels are heuristic rather than inferred chords: they are determined only by pitch interval above the local bass.
Real-piano (OOD) conditioning, as in the paper’s Table 2: estimate ± bootstrap SE, the SD over 1,000 song-cluster draws. Shape in dB; B+D (bass + drums) and Piano are HTDemucs energy shares in %. APA uses PCA-32. Ground-truth APA is omitted because the 476 pairs are not temporally aligned; FAD and Shape against ground truth are 0 by construction. Lower is better for FAD and Shape, higher for APA.
Differences are first system minus second system. Each cell is point difference [95% basic-bootstrap CI]; † marks an interval that excludes zero.
Each metric uses 476 held-out songs × five windows. The same 1,000 seed-42 song-cluster draws are shared across systems, retaining all five windows from each selected song. APA’s 33-pair aligned/misaligned reference, the full ground-truth FAD distribution, and the full ground-truth mean mel shape stay fixed. APA is computed in a 32-dimensional PCA space fitted to the aligned reference (no whitening; 89.8% of reference variance): its 171 reference windows cannot estimate a full 512-dimensional CLAP covariance. System rankings are unchanged for other k (e.g., 64).
These intervals are conditional on the frozen references and are not adjusted for testing many system pairs. HTDemucs energy shares are separator-dependent estimates.