krishnakalyan3/mml26-singing-synthesis-checkpoints

MML26 Singing Synthesis Checkpoints

Eight Basic Pitch checkpoints, including a MergeKit model soup selected on the Klangio validation set.

Best-checkpoint test metrics

Dataset	COnPOff_f1	COnP_f1	COn_f1	Best Epoch	Best Checkpoint Path	W&B Run	Runtime
Dataset 1 — syntheticdataset	0.1105	0.4595	0.6580	28	checkpoints/syntheticdataset-best-epoch28-COnPOff_f1-0.1105.ckpt	Not created (TensorBoard)	~790 s
Dataset 2 — syntheticdataset_soulx	0.2498	0.4773	0.6419	22	checkpoints/syntheticdataset-soulx-best-epoch22-COnPOff_f1-0.2498.ckpt	Not created (TensorBoard)	782 s
Dataset 3 — soulx_pseudo_student	0.2877	0.4707	0.6677	27	checkpoints/soulx-pseudo-student-best-epoch27-COnPOff_f1-0.2877.ckpt	Not created (TensorBoard)	734 s
Dataset 4 — syntheticdataset_aceopencpop	0.2137	0.3675	0.4842	16	checkpoints/syntheticdataset-aceopencpop-best-epoch16-COnPOff_f1-0.2137.ckpt	vkk8wmxv	898 s
Dataset 5 — syntheticdataset_aceopencpop_soulx	0.2424	0.4400	0.5933	27	checkpoints/syntheticdataset-aceopencpop-soulx-best-epoch27-COnPOff_f1-0.2424.ckpt	smau9uor	1582 s
Dataset 6 — Synthetic + ACE + SoulX	0.2388	0.4420	0.6236	25	checkpoints/syntheticdataset-all-three-best-epoch25-COnPOff_f1-0.2388.ckpt	9r5p4hoa	2228 s
Dataset 7 — audio augmentations (room_reverb, phone_codec, noisy_codec)	0.2414	0.4246	0.6644	75	checkpoints/best-epoch75-step003116.ckpt	yhns6pet	—
Dataset 8 — MergeKit (99% Dataset 7 + 1% Dataset 2)	0.2409	0.4199	0.6615	MergeKit	checkpoints/mergekit-dataset8-aug75-soulx-99-01-COnPOff_f1-0.2409.ckpt	Not created (MergeKit)	—

Evaluation context

• Datasets 1, 2, 4, 5, 6, 7, and 8 were evaluated on the same 35-song Klangio validation set.
• Dataset 3 used 28 teacher-pseudo-labeled Klangio songs for training and a fixed 7-song clean holdout for validation/test. Its metric is therefore not directly comparable to the 35-song metrics without accounting for the smaller evaluation set.
• All six used 30 epochs, 8-second chunks, batch size 32, Adam at 1e-4, onset/frame weights 18/9, BF16, and torch.compile(reduce-overhead).
• Dataset 4 contains 25.52 hours, 330 tracks, 192,086 pitched notes, and 30 ACE-OpenCpop synthetic voices. The derived training data retains the source dataset's CC BY-NC 4.0 terms.
• Dataset 5 is the 51.07-hour ACE-OpenCpop + SoulX union; Dataset 6 is the union of procedural, SoulX, and ACE-OpenCpop synthetic singing (1,130 tracks).
• Dataset 8 is a normalized linear merge made with MergeKit 0.1.4 from 99% Dataset 7 and 1% Dataset 2. Its decoding thresholds were selected on Klangio (onset=0.70, frame=0.01). These are validation scores, not held-out test scores. The reproducible merge config is mergekit/dataset8-merge.yml.
• metadata.json records SHA-256 hashes, exact checkpoint epochs, steps, and scores.

An earlier, separate SoulX experiment is available at W&B run ltqkp16d, but it produced a different checkpoint and is not the tracking run for Dataset 2 above.