# Third-party software and model notices

Olenji is an independently authored application. It uses public projects and public model artifacts through adapters; it does not incorporate FluidVoice's desktop UI or private enhancement service.

The application uses FastAPI, Starlette, Pydantic, Uvicorn, websockets, httpx, NumPy, SoundFile, sounddevice, huggingface-hub, Rich and jsonschema. Optional inference environments use faster-whisper/CTranslate2, transcribe.cpp, sherpa-onnx, ONNX Runtime, or the public FluidAudio Swift package. Windows installations include SoundCard for the loopback adapter. Native audio helpers use operating-system frameworks. FFmpeg is a separate prerequisite and its applicable license depends on the installed build. jsdom and Playwright are development-only browser-test dependencies; browsers and their automation runtimes are not bundled in the application wheel.

These dependencies are obtained from their normal package registries or pinned upstream sources. Their licenses are not replaced by Olenji's Apache-2.0 license. The release source and wheel do not bundle model weights, native inference binaries, FFmpeg, Python itself, or an entire optional runtime environment.

Model metadata records publisher URLs, artifact identities, expected license information, platform restrictions, and exact-versus-alternative parity provenance. `olenji models inspect ID` and `olenji models plan-install` expose this information before download. Installation manifests retain immutable source revisions and verified file digests where available.

Publisher conversions may use a different license from an upstream model. For example, the audited Cohere Core ML conversion declares GPL-3.0. Fluid Speech Pico is available through its public publisher-pinned archive, whose bytes and hashes were verified during this audit. The archive did not include a separate weight license, so the catalog reports `unresolved` and does not grant redistribution rights or bundle those weights. Apple assets are installed and managed through supported operating-system mechanisms rather than repackaged into Olenji.

The public speech fixture used for real inference verification is the `tests/jfk.flac` recording in OpenAI's Whisper repository: https://github.com/openai/whisper/blob/main/tests/jfk.flac. It is used only in the validation workspace; it is not bundled as user content or a fake transcript in the production application.

The bundled browser uses locally authored CSS, vanilla JavaScript and inline SVG symbols. It does not load remote fonts, analytics, icon CDNs or advertising scripts.