The best singing voice synthesis model trained on the Opencpop corpus scored 3.70 on a 5-point mean opinion score (MOS) scale, while the professional singer’s own recordings scored 4.51 (Wang et al., Opencpop paper, Interspeech 2022). In four other benchmark papers the best model landed between 3.81 and 4.01, against 4.30 to 4.54 for real recordings (see section 4). Humans are not perfect either: 24 adults singing Happy Birthday from memory missed their target notes by a median of 19 cents, about a fifth of a semitone (Mauch, Frieler and Dixon, JASA 2014), and the median listener only detects a mistuned vocal at around 55 cents (Larrouy-Maestri, Harrison and Müllensiefen, Behavior Research Methods 2019). Speech synthesis has reached parity on several benchmarks, as our synthetic speech MOS statistics roundup shows, but singing has not. We aggregated data from the Opencpop, M4Singer, GTSinger and ACE-Opencpop corpus papers, the Singing Voice Conversion Challenges of 2023 and 2025, the SVDD 2024 deepfake detection challenge, the SingMOS-Pro rating benchmark, and peer-reviewed singing research in JASA, Frontiers in Psychology, PLoS One and the Journal of Research in Music Education.
TL;DR
- The best Opencpop baseline scored 3.70 MOS vs 4.51 for the original recordings (Wang et al., Opencpop paper, Interspeech 2022).
- 24 adult singers hit notes with a median error of 19 cents (Mauch, Frieler and Dixon, JASA 2014).
- 28% of 50 occasional singers were classed as poor-pitch singers (Berkowska and Dalla Bella, Frontiers in Psychology 2013).
- Only 37% of 1,046 people were accurate pitch matchers under a strict 50-cent rule (Pfordresher and Demorest, JRME 2021).
- The median listener detects mistuning at a 54.989-cent shift (Larrouy-Maestri, Harrison and Müllensiefen, Behavior Research Methods 2019).
- Giving DiffSinger the real singer’s pitch curve lifted prosody MOS from 3.72 to 4.02, but audio quality only from 3.92 to 3.97 (Zhang et al., M4Singer, NeurIPS 2022).
- VISinger 2 fell from 3.96 to 3.02 MOS without its DSP synthesizer (Zhang et al., VISinger 2 paper 2022).
- CoMelSinger came within 0.24 naturalness points of recordings for unseen singers (Zhao et al., CoMelSinger 2026).
- ACE-Opencpop adds 128.9 hours of synthesized singing from 30 voices (Shi et al., ACE-Opencpop paper 2024).
- No system reached human-level naturalness in SVCC 2025, with top systems near 3.7 vs about 3.9 for real singing (SVCC 2025 analysis paper).
- SingMOS-Pro collected 44,247 quality ratings from 78 annotators (SingMOS-Pro benchmark paper 2026).
- The best SVDD 2024 detector reached a 1.65% equal error rate (Zhang et al., SVDD 2024 challenge paper).
1. The Human Baseline: How Accurately People Really Sing
Pitch accuracy is measured in cents: 100 cents equal one semitone, the step between two neighbouring piano keys. A median note error of 19 cents across 24 unaccompanied singers (Mauch, Frieler and Dixon, JASA 2014) is the realistic target for a synthetic singer, not zero. Even trained voices stray: professionals became less exact on intervals, not more, when they switched to operatic technique.
The spread across the population is wide. Only 24% of untrained participants were accurate pitch matchers in the same dataset where 37% of the full sample were (Pfordresher and Demorest, JRME 2021). Tempo explains much of the difference: in a classic study of 62 occasional singers, most of the 15 nonmusicians retested at a slower tempo matched professional accuracy, and only two kept making many pitch errors (Dalla Bella, Giguère and Peretz, JASA 2007).
| Metric | Value | Source |
|---|---|---|
| Median note error, 24 adults singing Happy Birthday without accompaniment | 19 cents | Mauch, Frieler and Dixon, JASA 2014 |
| Median absolute pitch drift over a roughly 50-second performance | 11 cents | Mauch, Frieler and Dixon, JASA 2014 |
| Mean interval deviation, 14 professional singers singing without operatic technique | 32.81 cents | Larrouy-Maestri and Morsomme, Logopedics Phoniatrics Vocology 2014 |
| Mean interval deviation, 63 occasional singers with no formal training | 50.83 cents | Larrouy-Maestri and Morsomme, Logopedics Phoniatrics Vocology 2014 |
| Mean interval deviation, the same professionals using operatic (lyrical) technique | 62.13 cents | Larrouy-Maestri and Morsomme, Logopedics Phoniatrics Vocology 2014 |
| Occasional singers classed as poor-pitch singers (2 standard deviations from group mean) | 28% (14 of 50) | Berkowska and Dalla Bella, Frontiers in Psychology 2013 |
| Accurate pitch matchers (90% or more of notes within 50 cents), combined sample of 1,046 | 37% | Pfordresher and Demorest, JRME 2021 |
| Singing classified as accurate when the criterion is loosened to 100 cents | 94% | Pfordresher and Demorest, JRME 2021 |
Most recent available data: Pfordresher and Demorest, Journal of Research in Music Education 2021. The poor-pitch share swings with the cut-off: on single sung pitches, 6% of the same 50 occasional singers fell beyond the 2 SD line but 58% fell beyond 50 cents, so any headline prevalence figure is only as good as its threshold.
2. Listener Tolerance: How Much Pitch Error Ears Forgive
Synthetic singers do not need to beat a tuner, they need to beat a listener. The median listener’s mistuning-detection ability matched a vocal pitch shift of 54.989 cents in the 333-person calibration of the Mistuning Perception Test, with 95% of listeners between 36.309 and 82.589 cents (Larrouy-Maestri, Harrison and Müllensiefen, Behavior Research Methods 2019). That is roughly half a semitone, well above the 19-cent median note error in section 1, although the two tests measure different things.
Ordinary listeners still judge pitch consistently. Lay judges’ scores tracked measured interval deviation almost as tightly as experts’ did, and they were stricter overall (Larrouy-Maestri et al., PLoS One 2015). For a synthesis model, that means a casual audience is a valid test panel for pitch errors, not just a panel of vocal coaches.
| Metric | Value | Source |
|---|---|---|
| Detection threshold, low vs high musical training (Gold-MSI training score 10 vs 40) | 65.173 vs 43.219 cents | Larrouy-Maestri, Harrison and Müllensiefen, Behavior Research Methods 2019 |
| Correct mistuning detection, smallest vs largest shift tested (10 to 100 cents, chance = 50%) | 51.2% to 81.5% | Larrouy-Maestri, Harrison and Müllensiefen, Behavior Research Methods 2019 |
| Correlation with measured interval deviation: 3 expert judges vs lay judges | about .83 vs about .79 | Larrouy-Maestri et al., PLoS One 2015 |
| Mean interval deviation across 166 Happy Birthday performances rated | 55.97 cents (range 10.10 to 383) | Larrouy-Maestri et al., PLoS One 2015 |
| Correlation between pitch discrimination and pitch matching, 632 volunteers | r(630) = -.31, p < .001 | Pfordresher and Demorest, JRME 2021 |
Most recent available data: Larrouy-Maestri, Harrison and Müllensiefen, 2019. The test shifted the whole vocal track by a constant amount, so it measures tolerance to a singer who is out of tune throughout, not to errors that vary from note to note, which the authors leave for future work.
3. Objective Pitch Metrics: What SVS Papers Measure
Singing voice synthesis (SVS) papers score pitch with F0 RMSE, the root mean square error between the generated and the real fundamental frequency (F0, the pitch curve). None of the SVS papers we reviewed reported that error in cents: the Opencpop, VISinger 2 and CSSinger tables give no unit at all, and ACE-Opencpop uses a logarithmic F0 scale. Human studies use cents, so a clean “machine vs human” pitch score does not exist yet.
What the numbers do show is ranking within a benchmark. On Opencpop, a Conformer model cut F0 RMSE from 29.99 to 26.72 compared with a Transformer baseline (Wang et al., Opencpop paper, Interspeech 2022), and VISinger 2 reached 26.0 with 25.7M parameters against CpopSing’s 137.5M (Zhang et al., VISinger 2 paper 2022). Semitone accuracy, reported as a percentage, is the closest metric to a human-style score, although the ACE-Opencpop paper does not spell out its formula.
| Metric | Value | Source |
|---|---|---|
| F0 RMSE on the Opencpop test set, Transformer vs Conformer baseline (unit not stated) | 29.99 vs 26.72 | Wang et al., Opencpop paper, Interspeech 2022 |
| F0 correlation with the real singer, Opencpop baselines | 0.893 to 0.904 | Wang et al., Opencpop paper, Interspeech 2022 |
| F0 RMSE at 22 kHz: VISinger 2 vs VISinger vs CpopSing | 26.0 vs 33.7 vs 28.5 | Zhang et al., VISinger 2 paper 2022 |
| Model size: VISinger 2 vs CpopSing | 25.7M vs 137.5M parameters | Zhang et al., VISinger 2 paper 2022 |
| Semitone accuracy of VISinger2: Mandarin ACE-Opencpop vs Japanese Kiritan corpus | 65.19% vs 49.33% | Shi et al., ACE-Opencpop paper 2024 |
| F0 RMSE, F0 correlation and voiced/unvoiced error of CSSinger-FS on Opencpop | 28.601, 0.919 and 0.107 | Cui et al., CSSinger 2024 |
| F0 RMSE on ACE-Opencpop: speech-language-model SVS vs TokSing vs XiaoiceSing | 62.79 vs 55.83 vs 71.67 | Zhao et al., speech language model SVS paper 2025 |
Voiced/unvoiced error tracks how often a model confuses voiced sound (the vocal folds vibrate, so a pitch exists) with unvoiced sound. The spread from 49.33% to 65.19% semitone accuracy for the same architecture suggests the score depends heavily on the corpus, so results from different corpora cannot be ranked against each other.
4. Naturalness: The MOS Gap Between Synthetic and Real Singing
MOS is the average rating that a listening panel gives on a 1 to 5 scale, and every paper here tested real recordings alongside the models. Across five benchmark papers the gap between the best model and real singing ranged from 0.45 to 0.81 MOS points (4.30 minus 3.85 on PopCS, 4.51 minus 3.70 on Opencpop). Speech synthesis has reported parity with recordings on several benchmarks; singing still trails by roughly half a point or more.
The newest zero-shot work narrows it. Zero-shot means the model sings in a voice it never heard during training. CoMelSinger, revised in April 2026, came within 0.24 naturalness points of real recordings for singers it had never heard (4.35 minus 4.11), the smallest gap in this set (Zhao et al., CoMelSinger 2026). Panels are small, though, often 10 to 20 listeners, so a 0.1-point lead is rarely decisive.
| Metric | Value | Source |
|---|---|---|
| MOS on Opencpop: best baseline (CpopSing) vs recordings | 3.70 vs 4.51 | Wang et al., Opencpop paper, Interspeech 2022 |
| MOS on PopCS: DiffSinger vs recordings | 3.85 vs 4.30 | Liu et al., DiffSinger, AAAI 2022 |
| Audio quality MOS on M4Singer: DiffSinger vs recordings | 3.92 vs 4.52 | Zhang et al., M4Singer, NeurIPS 2022 |
| MOS at 44.1 kHz: VISinger 2 vs recordings | 3.81 vs 4.32 | Zhang et al., VISinger 2 paper 2022 |
| Audio quality MOS on GTSinger: StyleSinger vs recordings | 4.01 vs 4.54 | Zhang et al., GTSinger, NeurIPS 2024 |
| Naturalness MOS, singers seen in training: CoMelSinger vs recordings | 4.02 vs 4.38 | Zhao et al., CoMelSinger 2026 |
| Naturalness MOS, unseen singers (zero-shot): CoMelSinger vs recordings | 4.11 vs 4.35 | Zhao et al., CoMelSinger 2026 |
Most recent available data for those corpora: the Opencpop, PopCS, M4Singer and VISinger 2 rows date from 2022, so they mark the baseline for each corpus, not the current state of the art. Wider adoption figures for AI music tools sit in our AI music generation statistics.
5. Where Pitch Matters: What Changes When the Model Gets F0 Right
The cleanest test hands the model the real singer’s pitch curve and measures what improves. When DiffSinger received the real F0, its prosody MOS rose by 0.30 points (3.72 to 4.02) while audio quality rose by only 0.05 (3.92 to 3.97) (Zhang et al., M4Singer, NeurIPS 2022). Prosody here means the coherence and naturalness of the melodic delivery. Better pitch closes part of that gap, but even with the real F0 both scores stay below the recordings (4.47 and 4.52), so pitch is one limit among several.
Dynamic techniques are the hardest. In the Singing Voice Conversion Challenge 2025, a conversion task rather than score-based synthesis, the styles the organizers describe as having strong dynamic components (breathy, glissando, vibrato) scored the lowest style accuracy, while mostly static ones such as falsetto scored the highest (SVCC 2025 analysis paper).
| Metric | Value | Source |
|---|---|---|
| F0 mean absolute error and voiced/unvoiced accuracy, DiffSinger predicted vs real F0 | 8.9990 vs 1.7061; 0.9406 vs 0.9856 | Zhang et al., M4Singer, NeurIPS 2022 |
| VISinger 2 MOS, full model vs without the DSP synthesizer | 3.96 vs 3.02 | Zhang et al., VISinger 2 paper 2022 |
| VISinger 2 MOS without the auxiliary mel-spectrum module | 3.85 | Zhang et al., VISinger 2 paper 2022 |
| SVCC 2025 style accuracy: breathy, glissando, vibrato | 37.3%, 42.6%, 43.9% | SVCC 2025 analysis paper |
| SVCC 2025 style accuracy: pharyngeal, mixed voice, falsetto | 44.8%, 48.3%, 48.8% | SVCC 2025 analysis paper |
The Opencpop authors reported that their baselines showed limited capability with rare pitches in the long tail of the training distribution, a limit that a single-singer corpus makes hard to escape.
6. Training Data: Hours of Singing Behind the Models
Singing models still learn from single- or double-digit hours of recorded singing. The Opencpop corpus that anchors much of the Mandarin SVS literature holds about 5.2 hours of singing from one professional female singer (Wang et al., Opencpop paper, Interspeech 2022). Recorded corpora such as Opencpop and M4Singer need note boundaries and musical scores annotated by hand, which is why real singing data grows slowly.
The workaround is synthetic data. ACE-Opencpop is derived from the Opencpop scores and synthesized with the ACE Studio singing synthesizer plus manual tuning of pitch contours, breaths and vibrato, producing 128.9 hours, and pre-training on it nudged VISinger2’s semitone accuracy on real Opencpop from 64.50% to 65.18% (Shi et al., ACE-Opencpop paper 2024). Pre-training on the real M4Singer corpus instead lowered it to 60.54%. ACE-Opencpop shares its songs with the target corpus, so this result says more about matching the target than about hours alone.
| Metric | Value | Source |
|---|---|---|
| Opencpop: 1 professional singer, 100 songs | about 5.2 hours, 3,756 utterances | Wang et al., Opencpop paper, Interspeech 2022 |
| M4Singer: 20 professional singers (10 male, 10 female), 700 songs | 29.77 hours, 20,942 utterances | Zhang et al., M4Singer, NeurIPS 2022 |
| GTSinger: 20 professional singers, 9 languages, 6 techniques | 80.59 hours singing plus 16.16 hours paired speech | Zhang et al., GTSinger, NeurIPS 2024 |
| ACE-Opencpop and ACE-KiSing synthesized corpora | 128.9 hours (30 voices) and 32.5 hours (34 voices) | Shi et al., ACE-Opencpop paper 2024 |
| VISinger2 semitone accuracy on Opencpop: no pre-training vs ACE-Opencpop vs M4Singer pre-training | 64.50% vs 65.18% vs 60.54% | Shi et al., ACE-Opencpop paper 2024 |
| VISinger2 on KiSing, trained alone vs with ACE-KiSing added: semitone accuracy and MOS | 68.12% to 72.51%; 3.58 to 3.83 | Shi et al., ACE-Opencpop paper 2024 |
| Real singing pooled from 8 public corpora for the CtrSVDD benchmark | 47.64 hours | Zang et al., CtrSVDD 2024 |
GTSinger’s nine languages are Chinese, English, Japanese, Korean, Russian, Spanish, French, German and Italian, and its six labelled techniques are mixed voice, falsetto, breathy, pharyngeal, vibrato and glissando, the same six styles scored in section 5.
7. Evaluation Methods and Benchmarks: How the Scores Are Produced
Small listening panels are the norm. Opencpop used 20 listeners per clip, the VISinger 2 test used ten listeners on 30 segments, and the ACE-Opencpop test used 30 listeners on 30 samples per system. Opencpop recordings scored 4.22 MOS in one test and 4.77 in another inside a single 2024 paper (Cui et al., CSSinger 2024), a 0.55-point swing larger than many of the gaps between models.
Automatic raters help at the system level but not per clip. The best MOS predictor on SingMOS-Pro reached a linear correlation of 0.88 when ranking systems, but only 0.58 on individual clips (SingMOS-Pro benchmark paper 2026). Detection is further along: on controlled data the top SVDD 2024 team cut equal error rate to 1.65% (Zhang et al., SVDD 2024 challenge paper), a topic covered in our audio deepfake forensics statistics.
| Metric | Value | Source |
|---|---|---|
| SingMOS-Pro size | 7,981 clips, 44,247 ratings, 78 annotators, 11.15 hours | SingMOS-Pro benchmark paper 2026 |
| Systems rated in SingMOS-Pro (synthesis, conversion, resynthesis, ground truth) | 141 (60, 17, 52 and 12) | SingMOS-Pro benchmark paper 2026 |
| Best automatic MOS predictor, linear correlation at system level vs clip level | 0.88 vs 0.58 | SingMOS-Pro benchmark paper 2026 |
| SVCC 2023 submissions and listeners | 26 submissions (incl. 2 baselines); 40 English and 319 Japanese listeners | Huang et al., SVCC 2023 paper |
| Spearman correlation of the UTMOS predictor with SVCC 2023 naturalness MOS (Task 1, Japanese listeners) | 0.77 | Huang et al., SVCC 2023 paper |
| SVCC 2025 naturalness, top systems vs real singing | about 3.7 vs about 3.9 | SVCC 2025 analysis paper |
| CtrSVDD real vs deepfake singing, 14 generation methods, 164 singer identities | 47.64 vs 260.34 hours | Zang et al., CtrSVDD 2024 |
| SVDD 2024 equal error rate: top team vs baselines; teams beating both baselines | 1.65% vs 11.16% and 12.03%; 37 of 47 | Zhang et al., SVDD 2024 challenge paper |
Context: in SVCC 2023 the top systems in both tasks were judged not significantly different from natural samples on naturalness, while in SVCC 2025, which scored singing style as well, no system reached human-level naturalness. The analysis concludes that modeling singing style still remains a challenge.
Summary: Singing Voice Synthesis Pitch Accuracy and Naturalness by the Numbers
| Metric | Value | Source |
|---|---|---|
| Median note error, 24 unaccompanied adult singers | 19 cents | Mauch, Frieler and Dixon, JASA 2014 |
| Accurate pitch matchers in a sample of 1,046 | 37% | Pfordresher and Demorest, JRME 2021 |
| Singing classified as accurate under a looser 100-cent criterion | 94% | Pfordresher and Demorest, JRME 2021 |
| Poor-pitch occasional singers (2 SD criterion) | 28% (14 of 50) | Berkowska and Dalla Bella, Frontiers in Psychology 2013 |
| Pitch shift matching the median listener’s detection ability | 54.989 cents | Larrouy-Maestri, Harrison and Müllensiefen, Behavior Research Methods 2019 |
| Best Opencpop baseline vs recordings, MOS | 3.70 vs 4.51 | Wang et al., Opencpop paper, Interspeech 2022 |
| VISinger 2 vs recordings at 44.1 kHz, MOS | 3.81 vs 4.32 | Zhang et al., VISinger 2 paper 2022 |
| CoMelSinger vs recordings, naturalness MOS for unseen singers | 4.11 vs 4.35 | Zhao et al., CoMelSinger 2026 |
| Prosody MOS gain vs audio quality MOS gain from real F0 (DiffSinger) | +0.30 vs +0.05 | Zhang et al., M4Singer, NeurIPS 2022 |
| VISinger 2 MOS without its DSP synthesizer | 3.02 (vs 3.96) | Zhang et al., VISinger 2 paper 2022 |
| VISinger2 semitone accuracy, ACE-Opencpop vs Kiritan | 65.19% vs 49.33% | Shi et al., ACE-Opencpop paper 2024 |
| SVCC 2025 style accuracy for vibrato | 43.9% | SVCC 2025 analysis paper |
| GTSinger corpus size | 80.59 hours, 20 singers, 9 languages | Zhang et al., GTSinger, NeurIPS 2024 |
| ACE-Opencpop synthesized corpus size | 128.9 hours, 30 voices | Shi et al., ACE-Opencpop paper 2024 |
| Opencpop recordings scored in two separate tests within one paper, MOS | 4.22 vs 4.77 | Cui et al., CSSinger 2024 |
| SingMOS-Pro ratings collected | 44,247 from 78 annotators | SingMOS-Pro benchmark paper 2026 |
| SVDD 2024 top equal error rate | 1.65% | Zhang et al., SVDD 2024 challenge paper |
Methodology and Sources
- Wang et al., Opencpop: A High-Quality Open Source Chinese Popular Song Corpus for Singing Voice Synthesis, Interspeech 2022
- Zhang et al., M4Singer: A Multi-Style, Multi-Singer and Musical Score Provided Mandarin Singing Corpus, NeurIPS 2022 Datasets and Benchmarks
- Liu et al., DiffSinger: Singing Voice Synthesis via Shallow Diffusion Mechanism, AAAI 2022
- Zhang et al., VISinger 2: High-Fidelity End-to-End Singing Voice Synthesis Enhanced by Digital Signal Processing Synthesizer, 2022
- Zhang et al., GTSinger: A Global Multi-Technique Singing Corpus with Realistic Music Scores for All Singing Tasks, NeurIPS 2024
- Shi et al., Singing Voice Data Scaling-up: An Introduction to ACE-Opencpop and ACE-KiSing, 2024
- Cui et al., CSSinger: End-to-End Chunkwise Streaming Singing Voice Synthesis System, 2024
- Zhao et al., Adapting Speech Language Model to Singing Voice Synthesis, December 2025
- Zhao et al., CoMelSinger: Discrete Token-Based Zero-Shot Singing Synthesis, revised April 2026
- Huang et al., The Singing Voice Conversion Challenge 2023
- An Extensive Analysis of the Singing Voice Conversion Challenge 2025 Evaluation Results, 2025
- Zhang et al., SVDD 2024: The Inaugural Singing Voice Deepfake Detection Challenge
- Zang et al., CtrSVDD: A Benchmark Dataset and Baseline Analysis for Controlled Singing Voice Deepfake Detection, 2024
- SingMOS-Pro: A Comprehensive Benchmark for Singing Quality Assessment, revised January 2026
- Mauch, Frieler and Dixon, Intonation in Unaccompanied Singing: Accuracy, Drift, and a Model of Reference Pitch Memory, JASA 2014
- Larrouy-Maestri and Morsomme, Criteria and Tools for Objectively Analysing the Vocal Accuracy of a Popular Song, Logopedics Phoniatrics Vocology 2014
- Larrouy-Maestri, Harrison and Müllensiefen, The Mistuning Perception Test, Behavior Research Methods 2019
- Larrouy-Maestri et al., Layman versus Professional Musician: Who Makes the Better Judge?, PLoS One 2015
- Berkowska and Dalla Bella, Uncovering Phenotypes of Poor-Pitch Singing: the Sung Performance Battery, Frontiers in Psychology 2013
- Pfordresher and Demorest, The Prevalence and Correlates of Accurate Singing, Journal of Research in Music Education 2021
- Dalla Bella, Giguère and Peretz, Singing Proficiency in the General Population, JASA 2007
- Data watch: MOS values are not comparable across papers, because listener panels (10 to 30 people), sample rates and instructions differ, and Opencpop recordings scored 4.22 and 4.77 in two tests in the CSSinger paper. F0 RMSE units vary (unstated in most papers, logarithmic in ACE-Opencpop), so F0 RMSE figures from different papers must not be compared directly. Sources conflict on corpus sizes: ACE-Opencpop is 128.9 hours in its own Table 1, about 130 hours in its introduction and 135 hours in the December 2025 speech-language-model paper; ACE-KiSing is listed with both 34 and 33 singers; Opencpop is about 5.2 hours in its own paper but 5.25 hours in the M4Singer and GTSinger comparison tables. CSSinger-FS F0 RMSE appears as 28.601 in one table and 28.101 in another. The SVDD 2024 equal error rates exclude the ACESinger method. Pfordresher and Demorest describe 50 cents as one semitone, while 50 cents is half a semitone. SVCC 2023 exact naturalness scores appear only in figures, so we report only the significance findings. VISinger 2, CSSinger, CoMelSinger, SingMOS-Pro and the speech-language-model paper are cited from arXiv versions and were scored by their own authors; CoMelSinger and SingMOS-Pro list no peer-reviewed venue, and the speech-language-model paper was submitted to a workshop, so final figures may differ. The human pitch-accuracy studies all predate 2023 and are the most recent available data.
Last updated: October 31, 2026. We update this roundup quarterly, and the next refresh is expected when the ICASSP 2027 proceedings and the next Singing Voice Conversion Challenge results are published.