RVC model catalog report: 7,733 published models
AIVoices grew from 678 to 7,733 published model records in three days. Verified language coverage also grew, but not fast enough to keep pace. This report shows where the catalog is deep, where the metadata is still thin, and what we are prioritizing next.
Four findings from the current catalog
11.4× the July 15 baseline
7,681 RVC-family records
922 models
Largest catalog share: 14.9%
Catalog growth outpaced language enrichment
The number of models with verified language information rose from 146 on July 15 to 922 on July 18—a 6.3× increase. But the complete catalog grew faster. Verified language coverage fell from 21.5% to 11.9% of published records.
This does not mean the remaining 6,811 models have no language. It means the available filename, repository, or reviewed metadata did not support a public language label confidently enough. AIVoices keeps those records out of language hubs instead of guessing.
Japanese leads the verified subset
Japanese accounts for 493 of the 922 models with verified language details. English follows with 175, while the two stored Spanish labels together account for 171. Small counts retain their stored labels here so this report does not silently merge regional or code values.
“Verified” describes support in the stored model details. It is not an audio-quality score and does not mean the model performs equally well across every accent, input voice, or use case.
RVC remains the catalog’s center of gravity
RVC and RVC v2 account for 7,681 of7,733 published model records. The current catalog therefore reflects community RVC distribution far more than the wider AI voice market.
This concentration is useful for focused RVC discovery, but it also means AIVoices cannot yet claim equivalent catalog depth for modern TTS or other voice-conversion frameworks.
The catalog is broader than its largest repositories
AIVoices now attributes published models to 589 creator profiles. The largest profile represents 14.9%of records, and the top three together represent 31.3%. That is still meaningful concentration, but no single source now defines most of the catalog.
A creator profile records source attribution. It does not by itself prove that the profile has been claimed or that every model was trained by the account owner.
Browse attributed creators →Most voice groupings still have one model
The snapshot contains 619 voice groupings connected to 887 model records. Of those groupings, 429 contain one model and 190 contain two or more. The multi-model groupings are the most useful for comparing separate creators, languages, and training runs without merging the individual downloads.
These are catalog relationships, not official character or rights owner groupings. Names awaiting review remain outside the indexable voice-page set.
Find voices with comparable models →The next bottleneck is understanding, not collecting
Raw catalog scale is no longer the only constraint. The larger opportunity is turning safe, source-linked records into useful discovery paths without inventing language, identity, franchise, or quality claims.
- Improve language coverage.Preserve unverified records in the full catalog while expanding exact, reviewable language evidence.
- Resolve more voice relationships.Group separate models only when character or voice evidence is strong enough to support the connection.
- Broaden framework coverage deliberately.Add TTS and other frameworks when AIVoices has real inventory, useful model details, and a clear product role.
How this snapshot was produced
We ran npm.cmd run seo:report-inventory once against the production catalog on July 18, 2026. The reporter uses seven aggregate SQL queries covering published model totals, frameworks, stored language states, franchise relationships, creator attribution, categories, and models per linked voice. It returns summary rows rather than downloading model records or files.
“Published” means the model record was in the published moderation state at capture time. It does not measure downloads, popularity, audio quality, model safety, legal permission, or current upstream availability. Public model pages apply additional source and listing rules. Counts can also change after corrections or removals.
The July 15 comparison uses the same inventory command and the dated baseline retained in our SEO audit. Percentages in this article are calculated from the stored snapshot rather than entered separately.