Models and languages

Reading a row

The catalogue is extracted from the 0.12.0 release source. It includes model package IDs, language declarations, voice modes, model-specific controls, licences and source links. Quantized packages are listed as variants when their declared capabilities match.

Documented languages means the source gives an explicit language list. Partial list establishes only the listed languages. Source claim preserves an upstream language-count claim without turning it into an exact list. Not established means the catalogue does not establish coverage. By variant means the checkpoints in that family have different language declarations.

Choose for your material

Filter by a language, then check the voice mode and delivery controls. Test difficult names, numbers and a representative paragraph. Compare samples with the same text, while recognizing that different voices and directions affect the result.

The table does not establish speed, required VRAM or identical quality across languages. Precision, backend, hardware, reference audio and source length all affect a run. API availability also depends on the configured account or deployment.

Local package requirements

A listed language can require extra runtime files. For Japanese, the standard small Kokoro GGUF omits UniDic. Use a fully bundled multilingual export and the MeCab library; see audio.cpp’s Kokoro setup.

Recognition versus timing

Speech recognition and alignment have separate language coverage. A transcript-only language is useful for text extraction but insufficient for a timed subtitle workflow without another timing route.