GitHub's build now runs the voice check on all three platforms: it installs Kokoro (English) and Piper (German) from scratch, and each one speaks the settings' sample sentence. All six checks passed on the first run, Windows included, so no fix was needed.
| Platform | Voice | Install | Sample sentence | Took | |
|---|---|---|---|---|---|
| Windows | Kokoro (en) | 26.8 s | 6.1 s of speech, 24 kHz | 21.4 s | ✓ |
| Piper (de) | 18.5 s | 5.5 s of speech, 22 kHz | 7.8 s | ✓ | |
| Linux | Kokoro (en) | 6.4 s | 6.1 s of speech, 24 kHz | 7.0 s | ✓ |
| Piper (de) | 1.7 s | 5.0 s of speech, 22 kHz | 3.8 s | ✓ | |
| macOS | Kokoro (en) | 8.6 s | 6.1 s of speech, 24 kHz | 9.4 s | ✓ |
| Piper (de) | 3.3 s | 5.6 s of speech, 22 kHz | 5.4 s | ✓ |
A clip passes when it's a 16-bit WAV more than 3 seconds long that isn't silent. Kokoro's clip is the same length and loudness on every platform (RMS 0.072), so it's the same model and the same speech everywhere. Windows is the slowest of the three: a Kokoro clip takes about 21 s there, against 7 to 9 s on Linux and macOS.
GitHub Actions run 42377357486 (build of main, 7 Nov 2026), jobs “voices (windows-latest)”, “voices (ubuntu-latest)”, “voices (macos-latest)”.