arXiv · 2025 · Preprint
Violeta et al. · → Paper · Demo: ? · Code: ✓
SVCC 2025 extends the Singing Voice Conversion Challenge series by shifting the target from singer identity conversion to singing style conversion, introducing two new tasks, a purpose-built dataset, and a large-scale evaluation across 26 systems from 7 teams.
Problem
Prior singing voice conversion challenges (VCC series and SVCC 2023) focused exclusively on converting the global timbre and singer identity of a voice. While top systems in SVCC 2023 approached ground-truth naturalness for that narrower task, singing style, including the dynamic features that define techniques such as vibrato, glissando, and breathy voice, has received no systematic evaluation. The challenge for 2025 is whether current systems can jointly convert both static identity features and the dynamic temporal patterns that characterize distinct singing styles.
Method
SVCC 2025 defines two tasks using a subset of the GTSinger corpus. Task 1 (in-domain singing style conversion) requires converting a seen singer’s voice from one style to another; all training styles are available but the test song is held out. Task 2 (zero-shot singing style conversion) requires the same style conversion for a singer whose data does not appear in training, testing the ability to disentangle singer identity from singing style. Seven styles are targeted: breathy, falsetto, mixed voice, pharyngeal, glissando, vibrato, and a neutral control style.
Two open-sourced baseline systems frame the evaluation space. Baseline 1 (Serenade) uses a flow-matching model conditioned on ContentVec, loudness, and MIDI for audio infilling, with a cyclic training strategy to disentangle source style. Baseline 2 (Vevo1.5) pairs an autoregressive language model with a flow-matching transformer: a chromagram-based melody tokenizer extracts coarse melody, the ARLM generates content-style tokens prompted by a style reference, and a flow-matching decoder produces the mel spectrogram conditioned on a timbre reference. Baseline 3 (NU-SVC) uses a flow-matching decoder adapted from a denoising diffusion architecture with ContentVec, log F0, and speaker-style token conditioning.
Submitted systems (S1-S7) build on these baselines with modifications including two-stage F0 and spectral feature prediction (S1), F0 fluctuation features (S2), dual HuBERT/Whisper content encoders (S3), Whisper-based style disentanglement with super-resolution post-processing (S4), residual style adapters (S5), and DPO or GRPO post-training of the ARLM (S6, S7).
Subjective evaluation uses crowd-sourced listening tests: MOS for naturalness, an XAB test for style similarity, and a four-point singer identity test with three reference samples per target singer. Objective evaluation uses the VERSA toolkit across 30-plus metrics, with Spearman rank correlation against subjective scores used to identify the most predictive measures.
Key Results
No submitted system achieved human-level naturalness. Top systems reached approximately 3.7 MOS naturalness compared to the ground truth at approximately 3.9 MOS. For singing style similarity, the gap was larger: top systems (S6, based on Vevo1.5 with DPO post-training) reached approximately 70% style accuracy, while ground truth was approximately 90%. For singer identity similarity, five systems showed no statistically significant difference from ground truth, which itself only reached approximately 70% due to the inherent difficulty of the evaluation setup.
Style-by-style analysis reveals where the difficulty lies. Breathy (37.3%), glissando (42.6%), and vibrato (43.9%) were the hardest styles to model accurately, while pharyngeal (44.8%), mixed voice (48.3%), and falsetto (48.8%) were comparatively more tractable. The easier styles share static spectral characteristics; the harder styles require accurate modeling of time-varying dynamics (noise envelope, F0 modulation trajectories).
Among objective metrics, neural MOS predictors (SHEETSSQA, SingMOS) correlated best with subjective naturalness (SRCC above 0.6), while speaker and singer embeddings had the highest correlation with identity and style similarity scores (SRCC above 0.75). Traditional signal-based metrics showed weaker correlation across all dimensions.
Novelty Assessment
The paper’s contribution is organizational and evaluative rather than architectural. SVCC 2025 is the first public challenge to define singing style conversion as a distinct task from singer identity conversion, and it provides a reproducible evaluation infrastructure: a purpose-built dataset from GTSinger, open-sourced baselines, and a large-scale crowd-sourced listening test with three reference samples per target identity. The shift from identity-only to joint identity-plus-style conversion is a meaningful problem reframing that exposes a gap prior challenges did not test. However, the individual submitted systems are all built on prior architectures; no fundamentally new model is introduced by the challenge organizers themselves.
Field Significance
Moderate — SVCC 2025 provides the first structured community benchmark for singing style conversion, establishing a reference dataset, task definitions, and evaluation protocol that subsequent work can build from. The challenge reveals that dynamic vocal features remain a hard open problem even as singer identity conversion has largely converged toward human-level scores, redirecting research attention toward temporal modeling in singing synthesis.
Claims
-
supports: Dynamic temporal features in singing (vibrato, glissando, breathy voice) are substantially harder to convert than static spectral features, and current systems cannot model them reliably.
Evidence: Style accuracy scores for vibrato (43.9%), glissando (42.6%), and breathy (37.3%) were markedly lower than those for falsetto (48.8%), mixed voice (48.3%), and pharyngeal (44.8%); no system approached human-level performance on the dynamic styles. (§4.1)
-
supports: Jointly converting singer identity and singing style simultaneously degrades naturalness compared to identity-only conversion.
Evidence: Top SVCC 2025 systems achieved approximately 3.7 MOS naturalness, below the approximately 3.9 MOS ground truth, whereas prior SVCC 2023 top systems reached human-level naturalness on the simpler identity-only task. (§4.1)
-
supports: Autoregressive language model architectures with flow-matching decoders are competitive with specialized diffusion systems for singing voice conversion.
Evidence: Systems based on Vevo1.5 (ARLM + flow-matching) achieved the highest singing style similarity scores in the challenge, outperforming pure diffusion-based systems using SeedVC and Serenade. (§4.1)
-
supports: Neural MOS prediction models show stronger correlation with subjective naturalness scores for singing voice than traditional signal-based metrics.
Evidence: SHEETSSQA and SingMOS achieved Spearman rank correlation coefficients above 0.6 with subjective naturalness scores; speaker and singer embeddings were the strongest predictors of identity and style similarity (SRCC above 0.75). (§4.2)
-
complicates: Singer identity similarity evaluation in challenges is confounded by evaluation design: even ground truth recordings achieve only approximately 70% similarity scores with a single-sample reference.
Evidence: Ground truth samples reached only approximately 70% singer identity similarity in SVCC 2025 listening tests; the challenge responded by presenting three reference samples per singer (instead of one in prior challenges) to reduce this floor effect. (§4.1)
Limitations and Open Questions
Warning
Objective metrics show weak-to-moderate correlation with subjective scores across most dimensions, with only neural MOS predictors and embedding-based similarity reaching SRCC above 0.6. Objective-only comparisons for singing style conversion should be treated with caution until better proxy metrics are established.
The challenge attracted only 7 participating teams (26 systems including ablations), limiting the diversity of architectural approaches evaluated. The dataset is derived exclusively from GTSinger, restricting style diversity to the 7 included categories and to the singers chosen for Task 1 and Task 2. The zero-shot task evaluation is further complicated by the inherent ambiguity of singer identity evaluation (GT achieving only approximately 70% similarity), which may inflate the apparent success of some zero-shot systems. Open questions include how to model the time-varying noise and F0 dynamics that characterize breathy, vibrato, and glissando styles, and whether objective metrics can be developed with sufficiently high correlation to replace or supplement crowd-sourced listening tests at scale.
Wiki Connections
- Voice Conversion — establishes the first community benchmark that extends singing voice conversion from identity to style, defining a new evaluation frontier within the VC task family.
- Singing Voice Synthesis — the challenge is entirely singing-domain, targeting seven singing-specific style dimensions (breathy, falsetto, mixed voice, pharyngeal, glissando, vibrato) that have no equivalent in spoken-voice conversion.
- Subjective Evaluation — conducts large-scale crowd-sourced listening tests with MOS, XAB style similarity, and four-point singer identity tests, including a methodological change (three reference samples vs. one) to reduce GT floor effects.
- Evaluation Metrics — reports Spearman rank correlation analysis across 30-plus VERSA metrics, identifying neural MOS predictors and speaker/singer embeddings as the strongest proxies for subjective scores.
- Flow Matching — several top-performing baselines and submissions use flow-matching decoders, including Serenade (audio infilling) and the Vevo1.5 flow-matching transformer, which outperformed pure diffusion alternatives in style similarity.
- Speaker Adaptation — Task 2 (zero-shot singing style conversion) directly tests adaptation to unseen singers, with five systems achieving statistically indistinguishable singer identity scores from ground truth.
- Vevo2 (Vevo2 / Vevo1.5) — Vevo1.5 serves as one of the two official baselines and the foundation for two of the top-performing submissions (S6, S7) in the challenge.