- Update
voices.json: addfa_IR-magnoliasis-mediumand update the files of
10 re-released voices - Reload a Piper voice whose files changed on disk. The loaded voice was
cached by name only, so uploading a voice again under the same name in the
web UI kept speaking with the old model, and a deleted voice kept being
spoken, until another voice was used or the server restarted - Make web UI voice uploads atomic. Piper uploads wrote the config and then
streamed the model straight to their final paths, so a request during a large
upload could load a half-written model, and an interrupted upload left a
broken voice behind. Both files are now written under hidden temporary names
and renamed into place once complete. OmniVoice uploads are filled in a
hidden directory that is renamed into place, and hidden directories are
skipped when scanning reference voices