airunner-tts-vendor

Vendored MeloTTS and OpenVoice forks used by the TTS pipeline in Capsize-Games/airunner (AI Runner). This is third-party code, kept in its own repository rather than a subdirectory of airunner-services specifically so its licensing and upstream provenance stay physically separate from that GPL-3.0 distribution's own code, and so upstream diffs stay legible.

Licensing

Both melo/ and openvoice/ are separately MIT-licensed by their respective upstream projects. See melo/LICENSE and openvoice/LICENSE for the exact upstream copyright notices; neither this package as a whole nor Capsize LLC's own integration changes relicense either subtree. This repository does not carry a top-level LICENSE file that would override either subtree's own.

Provenance

Extracted from Capsize-Games/airunner at commit b57cb8e3d (2026-09-17), as issue #2195, part of the repository-split tracker #2185. History for the moved files is preserved (git filter-repo).

Each subtree's own README documents its upstream project and whichever pin information issue #2190 could establish for it -- see melo/README.md and openvoice/README.md. Neither README names an exact upstream commit; both were vendored from the upstream project's main branch at the time of integration (issue #2051 in the source repository) without recording the specific commit. That is a real gap, not an oversight being papered over here: if an exact upstream fork point is ever needed (for issue #2113 in the source repository, the dependency/model license inventory, or for comparing against a specific upstream release), it is not currently determinable from this repository's history and would need to be reconstructed by diffing against the upstream projects directly.

Local modifications

Both forks were decoupled from the application they used to live beside (issue #2190, before this extraction): no import here reaches into airunner, airunner_common, or airunner_services. Where the fork previously needed something app-specific (a language enum, a model storage path, a cache directory, the application's GPU-memory-clearing helper), it now either defines its own equivalent (melo/language.py, openvoice/language.py) or exposes an injectable resolver/hook that the host process registers at startup (set_tts_model_root_resolver, set_tts_model_base_resolver, set_cache_base_resolver, set_memory_cleanup_hook in melo/runtime_support.py and melo/api.py) -- see each subtree's own README for the full list.

Do not reformat, restyle, or lint-fix the vendored code. The value of keeping it in its own repository is that upstream diffs stay readable; a formatting pass defeats that on day one.

Install

The base install has no torch dependency at all -- melo/api.py and openvoice/api.py need torch to import, but installing this package does not require it, since torch needs an explicit --index-url for your platform (CPU vs a specific CUDA version):

pip install -e .
pip install "airunner-tts-vendor[ml]" \
  --index-url https://download.pytorch.org/whl/cu129   # GPU
pip install "airunner-tts-vendor[ml]" \
  --index-url https://download.pytorch.org/whl/cpu     # CPU
# Plus, per language:
pip install -e ".[zh]"   # Chinese (MeloTTS/OpenVoice)
pip install -e ".[jp]"   # Japanese
pip install -e ".[kr]"   # Korean
pip install -e ".[tw]"   # Taiwanese Hokkien (g2pkk)
pip install -e ".[gruut]"  # German/Spanish/French phonemization support

No dependency on airunner, airunner-common, or airunner-services -- this package installs and imports standalone. airunner-services depends on the published version of this package; see runtimes/openvoice_model_manager.py and runtimes/openvoice_runtime_helpers.py in that repository for its public surface (melo.api, melo.runtime_support, openvoice.api, openvoice.se_extractor, openvoice.mel_processing) and the resolver registration that reconnects it to the application after installing this package.

Testing

This fork has no tests of its own in the source repository; the TTS synthesis tests that exercise it (test_tts_synthesize_functional.py, test_model_load_security.py) stay in airunner-services, since they need the full daemon and a real GPU/model to run meaningfully. CI here is therefore limited to import-level sanity checks: that both subtrees import cleanly with no application dependency present, and that the language-enum/resolver integration points this repository exposes behave as documented. Full synthesis coverage (and the bobross.wav reference-speaker fixture it uses) stays in airunner-services as an integration test against this published package.