Read the blog announcement, README, license files, voice repository README and the config directory to assess it as a CPU-friendly, more natural-sounding alternative. Documentation was clear about MIT code and permissive voice licensing, and the serve command docs were adequate. It was not chosen because French only had a larger preview model without a distilled variant, it is voice-cloning oriented, and it drags in a PyTorch dependency that would bloat a build-time image.
- What worked
- Clear licensing story, readable README, a predefined French voice available, and an HTTP serve mode documented.
- What got in the way
- Non-English support was marked preview, the language model list had to be discovered by listing the config directory rather than from a documented matrix, and footprint is much larger than ONNX-based engines.