NOTICE
======

TranscriptX third-party model and dataset notice (0.9.7 → 1.0).
Not legal advice. Identifiers follow documented Hub / upstream cards and in-tree profile `licence=` fields.
See also: docs/dev/trust_privacy_model_governance_1_0.md

This product may download or bundle the following categories of third-party
components when the corresponding install profile or feature is enabled:

1. spaCy English models (en_core_web_sm / md / lg / trf) — spaCy model licences
2. sentence-transformers embedding models (e.g. all-MiniLM-L6-v2, all-mpnet-base-v2)
   — typically Apache-2.0 on Hugging Face Hub
3. Optional Hugging Face text-classification profiles pinned in-tree
   (emotion / sentiment) — see profile `licence=` fields and Hub cards
4. NRCLex / NRC emotion lexicon terms (emotion lexical path)
5. NLTK / VADER data for default sentiment
6. Optional BERTopic / UMAP / HDBSCAN stack and their transitive licences
7. Optional SpeechBrain / ECAPA voice embedding models for speaker match
8. Optional user-selected Ollama models (licence depends on the model the user pulls)
9. Optional NVIDIA NeMo ASR weights (e.g. Parakeet TDT, Canary) when Theme K / `parakeet_docker`
   Run-in-app or Copy-command recipes are used — typically **CC-BY-4.0** on Hugging Face / NGC;
   user-triggered download on first run. See docs/recipes/parakeet/README.md.
10. Optional IBM Granite Speech weights (`ibm-granite/granite-speech-5.0-2b`,
    `ibm-granite/granite-speech-4.1-2b-plus`) when the `granite_mlx` provider or mlx-audio
    Copy command is used — **Apache-2.0**; user-triggered Hub download on first run.
    See docs/recipes/granite-speech/README.md.
11. Optional NVIDIA Nemotron 3.5 Transcribe weights
    (`nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B` and FP8/NVFP4 variants) when
    Theme K / `nemotron_vllm` Run-in-app or Copy-command recipes are used —
    **OpenMDW-1.1**; user-triggered download on first local vLLM/NIM serve.
    Default endpoint is loopback only. See docs/recipes/nemotron-transcribe/README.md.

TranscriptX itself does not claim ownership of these weights or lexicons.
Local-first operation: analysis runs on the user’s machine; model downloads
occur only when a feature that needs them is enabled.

Voice / speaker-identity embeddings are privacy-sensitive; see Settings → Speakers
and VOICE_PRIVACY_USER_NOTICE (privacy notice version voice_privacy_notice.v2).