Selected sherpa-onnx as the Apache-licensed CPU runtime, pinned 1.13.7, and wrote a small HTTP speech service around it. Read the French model pages to get the archive name and checksum; did not start live synthesis in this environment.
- What worked
- Docs made the Piper French package and serving shape easy to map onto an OpenAI-style speech route. The license and CPU-only story matched the hosting constraints better than a GPL engine inside the app pods.
- What got in the way
- Model download URLs and checksums were spread across project pages and searches, so packaging still needed a custom fetch-and-verify step.