A new check installs the local demo voices from scratch, the way the "Install" button in the demo settings does, and has each one speak the settings' sample sentence. It runs here by hand and in the GitHub build on every platform.
| Platform | Kokoro (English) | Piper (German) | Where |
|---|---|---|---|
| macOS, Apple Silicon | installs, speaks | installs, speaks | this Mac |
| Linux, ARM | installs (30 s), speaks | download stalled in Docker's network; not a Piper fault | Docker on this Mac |
| Linux, x86-64 | installs, speaks | installs, speaks | Docker, emulated x86 |
| Windows, x86-64 | first run once the change is on main | same | GitHub build |
On every platform tested, the speech recognition heard Kokoro's sample sentence word for word: "This is how the voice sounds in a demo. The card now shows the link, and one click opens the page."
On the Mac, with Obeya's home in a deep temporary folder, Kokoro failed: the pronunciation tool it uses for unknown words (espeak-ng) cannot read its data under a path longer than about 150 characters, and it ended the process. A normal home (~/.obeya) is far below that, but a long user name or a custom home could reach it. Kokoro now uses a copy of that data in the system's temporary folder when the path is too long.
✓ kokoro (en): installs: Kokoro, Model onnx-community/Kokoro-82M-v1.0-ONNX (about 790 MB), 37.8 s ✓ kokoro (en): speaks the sample sentence: 6.1 s of speech at 24000 Hz in 5.5 s, RMS 0.072, 5.5 s ✓ piper (de): installs: Piper, Piper-Stimme de_DE-thorsten-high (about 265 MB), 109.9 s ✓ piper (de): speaks the sample sentence: 5.5 s of speech at 22050 Hz in 6.2 s, RMS 0.176, 6.2 s
Kokoro's environment is larger on Linux (about 450 MB) than on the Mac (about 270 MB); with the model, the settings now say about 790 MB for English.