Nuvio Reshaped TV tester build
A test build for people helping to find bugs before changes reach Beta. Expect things to break.
Highlights
- Audio sync can load a pre-optimized ORT version of the same English speech model. Graph optimization is done before download; the app retains its current recognition settings, background loading worker and three-minute reuse period.
- The first ORT use downloads about 74 MB once on an unmetered connection, or from Settings. Existing ONNX files remain usable offline or if the upgrade download fails. Settings deletion removes both versions.
- Each new model file is verified by size and SHA-256 before being installed. Partial or corrupted downloads never replace completed files.
- The statistics panel reports ORT or ONNX so the actual loaded format is visible. A cached ONNX model continues until a fresh load or app restart.
- Native host tests showed faster model loading and matching recognized tokens on three speech clips, with token timestamp differences up to 40 ms. TV loading time, memory use and dialogue accuracy still need device testing; a sub-10-second load is not guaranteed.
Commits
- Prepare portable ORT speech model v1
Based on official NuvioTV 1.1.0-beta.5.