# Third-party runtime notes

The local voice path was extracted from the sibling `chrome-llm-poc` project.

- Pocket TTS model metadata and graphs are under `assets/models/pocket/`, including their bundled provenance and license files.
- Silero VAD and Pipecat Smart Turn v3.2 model graphs are under `assets/models/turn/`. `PROVENANCE.json` pins their source revisions and SHA-256 hashes; the bundled MIT and BSD-2-Clause license texts are `LICENSE.SILERO` and `LICENSE.SMART_TURN`.
- The local Whisper-compatible feature extractor follows a pinned Hugging Face Transformers preprocessing contract. Attribution is in `assets/models/turn/NOTICE.WHISPER_FEATURES`; the repository's existing full Apache-2.0 text is `assets/runtime/LICENSE.sentencepiece`.
- ONNX Runtime Web and SentencePiece runtime files and licenses are under `assets/runtime/`.
- `assets/voices/interviewer-male.f32` and `assets/voices/interviewer-female.f32` are derived from CMU ARCTIC BDL (male) and SLT (female) recordings. CMU permits commercial and non-commercial use subject to retaining its copyright/conditions, marking modifications, and preserving author names. The exact notice, modifications, speaker sources, prepared-reference provenance, and hashes are in [`app/assets/CMU_ARCTIC_VOICES_LICENSE.txt`](assets/CMU_ARCTIC_VOICES_LICENSE.txt). Source recordings are not included.
- `assets/voices/interview-placeholder.f32` remains only as an inactive development fallback. No source or reference recording is included.

Alternative production voices require their own rights records covering the intended synthesis, web distribution, and any commercial use. A model or software license does not itself authorize reuse of a real person's voice or identity. The CMU ARCTIC license independently governs the two bundled voice-conditioning files.

## Interviewer portraits

The WebP portraits under `app/assets/avatars/recruiter/`, `hiring-manager/`, and `technical-lead/` were generated for this project with OpenAI's built-in image generation tool on 2026-11-14. They depict fictional people, are not based on supplied photographs, and are used as simulated interviewer personas. Their generation prompts requested no public-figure resemblance, logos, text, or watermark.

The matching mouth textures were generated from those fictional portraits using the built-in image generator. Only the small, aligned speech atlases ship with the app.

Vincent's portrait under `app/assets/avatars/vincent/` comes from the owner's supplied `slack vincent.png`. On 2026-11-15 the owner explicitly requested, “Use me as the default interviewer,” and authorized image generation for speaking frames. The built-in image generator produced the neutral, non-smiling portrait and its mouth textures. After the owner requested clearer speech and blinks behind the glasses, the generator produced new speaking poses and a closed-eyelid texture from the portrait references. The meeting room identifies Vincent as a simulated interviewer. This request authorizes the local app implementation. See [the preparation record and exact prompts](../docs/vincent-interviewer.md).

Playback analysis uses three unmodified modules from [Beer Digital's lipsync-engine, revision d4f1d4a](https://github.com/Amoner/lipsync-engine/tree/d4f1d4aaf85be9bf8a6e2264f828dd8cab019a92/src). The source and its MIT license are bundled in `app/vendor/lipsync/`. The app uses the frequency analyzer on a local Web Audio tap, not the library's streaming player or remote demos.

Before a public or commercial release using a real person's photograph, record permission covering that distribution and use.

## Interviewer asset restrictions

The [interviewer asset notice](assets/INTERVIEWER_ASSETS_NOTICE.txt) accompanies the portrait assets. Voice and portrait metadata link to the public Terms of Service. The notice covers only applicable rights in interviewer assets and identity; it does not replace or narrow the separate licenses for model graphs, ONNX Runtime, SentencePiece, feature extraction, or bundled open-source code. Your own answers and ordinary sharing of text coaching reports remain outside these asset restrictions. The notice and metadata do not add encryption, watermarking, access control, or a technical guarantee against copying.
