ZONOS3-Mini / PROVENANCE.md
fffiloni's picture
v0.3.0: refine quiet-wave display and trim state; record live validation
319d860 verified
|
Raw History Blame Contribute Delete
3.57 kB
# Source and dependency provenance
- Model: https://huggingface.co/Zyphra/ZONOS3-Mini/tree/eaac314642440e8f1f497f1505ca376f8e618f2a
- Official source: https://github.com/Zyphra/ZONOS3-Mini/tree/b820f6da278862d7c2db3e628d21cbae85983720
- Vendored subtree: `model/src/zonos3/` → `vendor/zonos3/`. Source, language data and registries are included; model weights are not.
- Root `LICENSE` and `NOTICE` from that source are retained as `vendor/ZYPHRA_LICENSE` and `vendor/ZYPHRA_NOTICE`; its third-party license files are retained in `vendor/third_party/licenses/`.
- The application is a new Space around the official Python runtime, not a fork of a deployed community Space. The visual references are the Zyphra presentation captures supplied for this project.
The only modified upstream source module is `vendor/zonos3/utils/release.py`: `ZONOS3_RELEASE_REVISION` is passed to model downloads for the configured release repository. This prevents mixed snapshots. The exact diff is in `patches/pin-model-revision.patch`; hashes of all 727 files in the subtree are recorded in `vendor-manifest.json`. None of that subtree's files was omitted. No inference loop, model layer, codec math or normalization behavior was rewritten. Separate steering assets are not included; this application does not enable steering. The two small FAR files inside the vendored normalizer are original language-rule data, not model weights or local caches.
`requirements.txt` pins the application stack. Gradio 6.30.0 requires `huggingface-hub>=1.16`, whereas the official source's Transformers 4.x pin requires an older Hub package. This integration therefore uses **Transformers 5.22.2 / huggingface-hub 1.31.0**, with import and dependency-resolution checks. The used MECT path is implemented in the official source; other speaker architectures imported by its module are not instantiated. The first HF build also revealed a preinstalled datasets 5.0.0 constraint of Hub <2.0. Hub 1.31.0 satisfies both Transformers and that base-image constraint. Full checkpoint loading and short inference runs subsequently passed on the deployed ZeroGPU Space.
Torch and torchaudio are both 2.9.1, a version supported by the reviewed ZeroGPU documentation. Gradio is pinned in both requirements and README metadata. The `spaces` SDK is pinned to 0.51.3. The SDK release path was inspected, but the application does not call its private APIs. Only `spaces.GPU`, `gr.Progress`, `gr.Request`, documented Gradio events and the public HTML component `props/watch/trigger` interface are used.
Official references reviewed:
- https://huggingface.co/docs/hub/spaces-zerogpu
- https://huggingface.co/docs/hub/spaces-config-reference
- https://www.gradio.app/docs/gradio/html
- The pinned Gradio 6.30.0 and spaces 0.51.3 package sources.
System fonts are used; no remote font or font binary is required. No weights, reference recordings, generated audio, environment, cache, token or local development path belongs in the distributable. `MANIFEST.sha256` covers packaged files; it intentionally does not hash itself.
Word alignment model: https://huggingface.co/Qwen/Qwen3-ForcedAligner-0.6B-hf/tree/c07281df297b9905d24a508279258cccf987a064 (Qwen team, Apache 2.0). The existing pinned Transformers 5.22.2 provides the public prepare_forced_aligner_inputs / decode_forced_alignment methods and AutoModelForTokenClassification implementation. No remote model code, NeMo runtime, vLLM server, private Gradio state or extra kernels are used. The HTML audio component uses the documented upload(file) helper.