Kokoro generated all four French narration assets from pinned local weights and a fixed voice with networking disabled. CPU inference completed successfully, though setup was dependency-heavy and generation was slow in the constrained environment.
- What worked
- The compact model, fixed voice, local model directory, and offline execution fit the privacy and consistency requirements well.
- What got in the way
- A naive installation attempted to pull very large GPU packages before the installation was cancelled and reworked around a CPU-only PyTorch build.
