AI song covers from reference vocals, without training · TelkNet

Voice & singingGPU workloadYingMusic-SVCSource Weights PaperHQ-SVCWeights Source PaperSoulX-Singer-SVCSource Weights PaperSeed-VCSource Weights PaperVevo2-FMSource Weights PaperSaMoye-SVC · expand_train_jxbSource Weights PaperkNN-SVCSource Weights Paper

Reference voice cover

Upload the original song and a target singer reference to change the lead voice without training. YingMusic-SVC / HQ-SVC / SoulX-Singer-SVC / Seed-VC / Vevo2-FM / SaMoye-SVC / kNN-SVC process the same inputs and return separate covers.

112 credits/useAudio file inputMax 100 MBEstimated 300-7200 seconds

Input

An original song or isolated vocals, plus a target singer reference.

Audio formats

mp3wavflacoggm4aaac

Output

Vocal WAV files for the selected models; song mode also includes accompaniment mixes. MP3 previews, original WAV and ZIP downloads.

Use case

Changing the lead voice from reference audio and comparing seven model outputs.

07 / Models

Seven models · individual metrics

No training

Demo: Far far away, full source 240.98 seconds, one target singer, seed 42 and pitch shift 0. kNN-SVC uses the full reference; the others share a 25-second excerpt. Voice similarity is measured against that same excerpt for every model. These measurements belong to this demo.

YingMusic-SVC

GitHub HF Paper

ASR text differenceBest in metric: 6.8% ↓
Median pitch error: 7.27cent ↓
Extra sound during pauses: 2.5s ↓
Reference timbre similarityBest in metric: 0.978 ↑
Predicted singing naturalness: 3.59 ↑
Predicted perceptual quality: 7.707 ↑

HQ-SVC

GitHub HF Paper

ASR text difference: 15.4% ↓
Median pitch errorBest in metric: 4.40cent ↓
Extra sound during pausesBest in metric: 0.0s ↓
Reference timbre similarity: 0.948 ↑
Predicted singing naturalness: 3.60 ↑
Predicted perceptual qualityBest in metric: 7.890 ↑

SoulX-Singer-SVC

GitHub HF Paper

ASR text difference: 10.2% ↓
Median pitch error: 7.44cent ↓
Extra sound during pauses: 12.6s ↓
Reference timbre similarity: 0.977 ↑
Predicted singing naturalness: 3.80 ↑
Predicted perceptual quality: 7.529 ↑

Seed-VC

GitHub HF Paper

ASR text difference: 7.4% ↓
Median pitch error: 8.36cent ↓
Extra sound during pauses: 5.5s ↓
Reference timbre similarity: 0.973 ↑
Predicted singing naturalness: 3.53 ↑
Predicted perceptual quality: 7.582 ↑

Vevo2-FM

GitHub HF Paper

ASR text difference: 27.1% ↓
Median pitch error: 38.66cent ↓
Extra sound during pauses: 3.1s ↓
Reference timbre similarity: 0.974 ↑
Predicted singing naturalness: 3.51 ↑
Predicted perceptual quality: 7.276 ↑

SaMoye-SVC · expand_train_jxb

GitHub HF Paper

ASR text difference: 14.2% ↓
Median pitch error: 11.47cent ↓
Extra sound during pauses: 1.2s ↓
Reference timbre similarity: 0.948 ↑
Predicted singing naturalnessBest in metric: 4.04 ↑
Predicted perceptual quality: 7.888 ↑

kNN-SVC

GitHub Weights Paper

ASR text difference: 20.3% ↓
Median pitch error: 5.35cent ↓
Extra sound during pausesBest in metric: 0.0s ↓
Reference timbre similarity: 0.962 ↑
Predicted singing naturalness: 3.61 ↑
Predicted perceptual quality: 7.804 ↑

Highlights the best value in this group according to each metric’s arrow, including ties; this is not an overall audio-quality ranking.

Lower ↓ or higher ↑ is better. Transcript difference compares Whisper-medium recognition with the source, not manually checked lyric errors. Timbre similarity is Resemblyzer embedding cosine similarity, not identity proof. SingMOS-Pro naturalness and AudioBox PQ are automatic predictions, not human MOS or the paper tables. Leading one metric on this song does not establish an overall winner across songs.

Papers and score sources

Authors’ results are shown separately by dataset and setting. They do not form a combined ranking of the seven models.

YingMusic-SVC · Table 2 HQ-SVC · Table 1 SoulX-Singer-SVC · Table 4 Seed-VC · EVAL.md Measurements and input verification

Published evidence

Papers and official evaluations

Authors’ results are shown separately by dataset and setting. They do not form a combined ranking of the seven models.

Only records with verified public inference implementations and matching weights are shown; algorithms requiring no weights are also eligible. Closed services, unreleased weights and unmatched experiment versions are excluded. This is not an all-model leaderboard.

YingMusic-SVC

arXiv:2512.04793v1 · Table 2

Evaluation setting

YingMusic · Table 2

Choose a dataset to view all metrics for each model. Rows follow the official rank or selected metric. Bold marks each column’s best comparable value, including ties. Different evaluation protocols are ranked separately; missing scores are shown as -.

Deployable entries 9 / source rows 15

Not ranked: 6

A complete public inference package for these experiments has not been verified. Other weights with the same model name do not qualify.

Ours-CPT · Setting 1: GT Leading · Ours-SFT · Setting 1: GT Leading · Ours-CPT · Setting 2: Mix Vocal · Ours-SFT · Setting 2: Mix Vocal · Ours-CPT · Setting 3: Ours Leading · Ours-SFT · Setting 3: Ours Leading

Test settingRank by metric

YingMusic · Table 2
Metric rank	Model	Test setting	SPK-SIM ↑	CER(%) ↓	CE ↑	CU ↑	LogF0PCC(%) ↑	CMOS ↑	SMOS ↑
1	Ours-Full Public weights · Inference code YingMusic-SVC-full.pt Inference code Weight license (repository declaration): cc-by-nc-4.0	Setting 1: GT Leading	0.800	9.26	5.86	6.56	98.12	4.26	3.08
2	Seed-VC ( Liu, 2024 ) Public weights · Inference code DiT_seed_v2_uvit_whisper_base_f0_44k_bigvgan_pruned_ft_ema_v2.pth Inference code Evaluation settings Weight license (repository declaration): gpl-3.0	Setting 1: GT Leading	0.801	10.89	5.69	6.39	98.29	3.98	3.09
3	FreeSVC ( Ferreira et al., 2025 ) Public weights · Inference code G_00014_0225000.pth spin.ckpt rmvpe.pt Inference code Evaluation settings Weight license (repository declaration): cc-by-nc-sa-4.0	Setting 1: GT Leading	0.648	20.70	4.72	5.55	86.56	2.47	1.83

Original records (9)

arXiv:2512.04793v1 · Table 2 ↗

Models, parameters, and sources

YingMusic-SVCHQ-SVCSoulX-Singer-SVCSeed-VCVevo2-FMSaMoye-SVC · expand_train_jxbkNN-SVC

Use cases

Cover an existing song with a new lead voice while retaining its original accompaniment and backing vocals in the mix.
Convert isolated vocals, enabling harmony separation or dereverb as the recording needs.
Compare diction, sustained notes, breaths and high notes across seven versions.

Workflow

1Upload the source and select its type. Full songs get accompaniment and harmony separation, then lead-vocal dereverb. Isolated vocals have separate switches for both cleanup steps.
2You can upload a complete song by one singer. After vocal cleanup, kNN-SVC uses the full reference; the other selected models share one continuous excerpt of up to 25 seconds. Vocal mode expects a dry solo voice. The source song is converted in full.
3Choose one or more models using the same song and voice reference. Select Start cover. Preview each converted vocal, or the mix in song mode. Previews are 192 kbps MP3; individual and ZIP downloads retain the original WAVs. Keep your place when switching models or vocals/mix. Playback switches when the next audio is ready. Drag the timeline and release to seek.

What to check

Check that the reference features the same lead singer throughout. Avoid choirs, duets and medleys; this tool does not verify singer identity.
Clean, dry solo vocals need not be separated or dereverberated again. Use full-song mode if accompaniment remains.
Separation residue can affect timbre and diction. Check choruses, high and sustained notes, and gaps between phrases. Diagnostic scores are not an overall quality ranking.

Models in this field

6 entries

Includes public test results for models used here and elsewhere. The complete source table is shown by default; conditions and metrics can also be selected for comparison. Scores from different test conditions are not directly comparable.

Public weights do not establish a fully open-source license. Expand the data for declared weight licenses and access conditions. Undeclared licenses are marked separately.

BenchmarkRanking category

The selected model has no verified matching public result and has not been ranked.

SoulX-Singer paper SVC table

Deployable entries 6 / source rows 10

Not ranked: 4

A complete public inference package for these experiments has not been verified. Other weights with the same model name do not qualify.

GroundTruth · GMO-SVS Dataset · Vevo-SVC · GMO-SVS Dataset · GroundTruth · SoulX-Singer-Eval Dataset · Vevo-SVC · SoulX-Singer-Eval Dataset

Only rows using the stated protocol are shown.

Test settingRank by metric

SoulX-Singer paper SVC table
Metric rank	Model	Test setting	GMO-SVS Dataset · Chinese					GMO-SVS Dataset · English
Metric rank	Model	Test setting	WER ↓	SIM ↑	FFE ↓	SingMOS ↑	Sheet ↑	WER ↓	SIM ↑	FFE ↓	SingMOS ↑	Sheet ↑
1	SoulX-Singer (Melody) Public weights · Inference code Soul-AILab/SoulX-Singer@40493ad90286056c7a9095035164434a79daa8c9 Inference code Evaluation settings Weight license (repository declaration): apache-2.0	GMO-SVS Dataset	0.065	0.897	0.044	4.458	4.110	0.151	0.918	0.036	4.323	3.751
2	SoulX-Singer-SVC Public weights · Inference code Soul-AILab/SoulX-Singer@40493ad90286056c7a9095035164434a79daa8c9 Inference code Evaluation settings Weight license (repository declaration): apache-2.0	GMO-SVS Dataset	0.107	0.904	0.036	4.419	4.029	0.267	0.919	0.030	4.314	3.666
3	YingMusic-SVC Public weights · Inference code YingMusic-SVC-full.pt Inference code Evaluation settings Weight license (repository declaration): cc-by-nc-4.0	GMO-SVS Dataset	0.097	0.908	0.057	4.399	3.974	0.243	0.923	0.046	4.337	3.708

Original records (6)

SoulX-Singer paper SVC table ↗Checked: 2026-10-25

All evaluation sources (1)

Datasets, versions and conditions remain separate, including sources with no eligible ranking rows.

SoulX-Singer paper SVC table ↗
Deployable entries 6 / source rows 10
Not ranked: 4
A complete public inference package for these experiments has not been verified. Other weights with the same model name do not qualify.
GroundTruth · GMO-SVS Dataset · Vevo-SVC · GMO-SVS Dataset · GroundTruth · SoulX-Singer-Eval Dataset · Vevo-SVC · SoulX-Singer-Eval Dataset

All original results and sources

Scores apply to the listed dataset, metric and model version; results from different tests are not directly comparable.

SoulX-Singer paper SVC table

Deployable entries 6 / source rows 10

Not ranked: 4

A complete public inference package for these experiments has not been verified. Other weights with the same model name do not qualify.

GroundTruth · GMO-SVS Dataset · Vevo-SVC · GMO-SVS Dataset · GroundTruth · SoulX-Singer-Eval Dataset · Vevo-SVC · SoulX-Singer-Eval Dataset

Only rows using the stated protocol are shown.

Test settingRank by metric

SoulX-Singer paper SVC table
Metric rank	Model	Test setting	GMO-SVS Dataset · Chinese					GMO-SVS Dataset · English
Metric rank	Model	Test setting	WER ↓	SIM ↑	FFE ↓	SingMOS ↑	Sheet ↑	WER ↓	SIM ↑	FFE ↓	SingMOS ↑	Sheet ↑
1	SoulX-Singer (Melody) Public weights · Inference code Soul-AILab/SoulX-Singer@40493ad90286056c7a9095035164434a79daa8c9 Inference code Evaluation settings Weight license (repository declaration): apache-2.0	GMO-SVS Dataset	0.065	0.897	0.044	4.458	4.110	0.151	0.918	0.036	4.323	3.751
2	SoulX-Singer-SVC Public weights · Inference code Soul-AILab/SoulX-Singer@40493ad90286056c7a9095035164434a79daa8c9 Inference code Evaluation settings Weight license (repository declaration): apache-2.0	GMO-SVS Dataset	0.107	0.904	0.036	4.419	4.029	0.267	0.919	0.030	4.314	3.666
3	YingMusic-SVC Public weights · Inference code YingMusic-SVC-full.pt Inference code Evaluation settings Weight license (repository declaration): cc-by-nc-4.0	GMO-SVS Dataset	0.097	0.908	0.057	4.399	3.974	0.243	0.923	0.046	4.337	3.708

Original records (6)

SoulX-Singer paper SVC table ↗Checked: 2026-10-25

Reference voice cover

Seven models · individual metrics

YingMusic-SVC

HQ-SVC

SoulX-Singer-SVC

Seed-VC

Vevo2-FM

SaMoye-SVC · expand_train_jxb

kNN-SVC

Papers and official evaluations

YingMusic-SVC

YingMusic · Table 2

Models, parameters, and sources

YingMusic-SVC

HQ-SVC

SoulX-Singer-SVC

Seed-VC

Vevo2-FM

SaMoye-SVC · expand_train_jxb

kNN-SVC

Official / repositories

Papers / technical notes

Use cases

Workflow

What to check

Models in this field

SoulX-Singer paper SVC table

SoulX-Singer paper SVC table

Quick feedback