VITS
The 2021 end-to-end TTS breakthrough: conditional VAE plus adversarial training with a stochastic duration predictor.
About
VITS is the 2021 research model that showed end-to-end TTS could beat two-stage pipelines at their own game. Proposed by Jaehyeon Kim, Jungil Kong and Juhee Son, it combines a conditional variational autoencoder (with normalizing flows) and adversarial training with a stochastic duration predictor, so one text input can be spoken many ways with different pitch and rhythm. On the LJ Speech dataset, the authors reported a mean opinion score on par with ground-truth recordings and above the best public TTS systems of the time.
The paper (arXiv 2106.06103, submitted June 11, 2021) became a foundation for a generation of open TTS work — MeloTTS, for example, builds directly on the VITS lineage. The official implementation is public for research use, with pretrained models and a Colab demo. All quality claims are author-reported; there is no independent benchmark. Check the repo's license terms before any commercial use.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
Author-reported MOS on LJ Speech: on par with ground truth, above best public TTS systems of 2021; no independent verification
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 8/10
Scored 8/10 (TIER verified, metrics-v1). Evidence: published paper (arXiv 2106.06103, June 11, 2021) with MOS evaluation; official public implementation with pretrained models; foundational influence on later open TTS (VITS2, Bert-VITS2, MeloTTS). Strengths: peer-reviewed-style results, open code, wide community adoption. Docked for: (1) all quality claims are author-reported with no independent replication; (2) repo is a research artifact with quiet maintenance; (3) license terms not clearly documented — verify before commercial use. License: unverified — no standard license file located in official repo.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 9.0/10
- Maintenance · recency, releases, issue load
- 4.0/10
- Openness · license + docs
- 8.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · Paper arXiv 2106.06103, submitted June 11, 2021 (Kim, Kong, Son)
- · Conditional VAE + normalizing flows + adversarial training + stochastic duration predictor
- · Author-reported MOS on LJ Speech on par with ground truth; foundation for VITS2/MeloTTS lineage
- · No logo asset; logoUrl omitted