VISinger 2
End-to-end singing synthesis at 44.1 kHz without the glitches: a VITS-style model with a DSP synthesizer that splits harmonics from noise before HiFi-GAN ever sees them.
About
VISinger 2 (arXiv 2211.02903, November 2022) fixes the two problems that plagued end-to-end singing synthesis: phase glitches from separately-trained two-stage pipelines, and weak modeling at higher sampling rates. It borrows from differentiable DSP: a harmonic synthesizer and a noise synthesizer generate periodic and aperiodic signals from the latent representation, and those feed a HiFi-GAN decoder as conditional inputs, so the latent space only has to carry amplitude information, not phase. Built on the VISinger/VITS lineage, it trains for 500k steps on the Opencpop dataset at 44.1 kHz, with a flow-based variant added in April 2023 for a more flexible prior. The reference implementation (ifgcguitarclub/visinger2) ships training and inference scripts plus pre-trained weights via Google Drive and demo audio. The paper reports it outperforming the original VISinger and two-stage baselines in listening tests (author-reported). One caveat: the repo carries no license file, so commercial terms are unverified.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- Reported to outperform VISinger and two-stage SVS baselines in MOS tests (author-reported)
- Ablations on the DSP synthesizer design and the flow-based prior (author-reported)
- No independent replication found
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 7/10
Scored 7/10 (TIER verified, metrics-v1). Evidence: published paper with full method detail, a complete training recipe, and pre-trained Opencpop weights anyone can download. Strengths: the DSP-synthesizer idea is a genuinely clever fix for phase glitches in end-to-end SVS, and the repo includes everything needed to reproduce training. Docked for: (1) the repo has no license file, so commercial use terms are unverified; (2) no pushes since 2023 and all figures are author-reported; (3) weights live on Google Drive, not a stable artifact host. License: none declared, verify before commercial use.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 6.0/10
- Maintenance · recency, releases, issue load
- 4.0/10
- Openness · license + docs
- 6.0/10
- Manual review · sanity read + red-flag check
- 7.0/10
- · arXiv 2211.02903, November 2022; DSP-synthesizer-enhanced end-to-end SVS at 44.1 kHz
- · Reference impl ifgcguitarclub/visinger2; pre-trained Opencpop weights via Google Drive
- · 500k training steps; flow-based prior variant added April 2023
- · Repo static since 2023; release date 2022-11-01 month-level approximation; slug 'visinger' used because 'visinger2' was taken