You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Word-level timelines are generated on device with ONNX Runtime: CTC for English and a separate CIF monotonic aligner for Chinese. Inference stays local; no audio is uploaded.
Model packs move to android-models-v0.1.0-beta.2, downloaded and verified per file by SHA-256; a matching installed pack is reused without re-downloading.
Generation jobs are serialised and stop on track change or when the switch is off; sessions are released afterwards, keeping at most one language model resident.
Authored word timing is never overwritten; lines that cannot be aligned keep their line timing and are never padded with an even split.
Japanese stays unavailable until the on-device G2P (OpenJTalk + NAIST dictionary) ships.
Player and visuals (from the contributor branch)
Rebuilt now-playing screen with Compose shared-element / shared-bounds one-shot transitions and a landscape layout rebuild.
Kyant0 Backdrop liquid-glass material, liquid bottom bar and transformed backdrops, plus a Mei mesh background, smooth corners and a display corner.
Haptics, a variable-refresh-rate frame controller, a system animation-scale bridge and a volume-key bridge.
NetEase WebView QR login, a song-recognition overlay and a dedicated settings detail page.
New MeloXCrashRecorder, a network retry interceptor, playback-unavailable notices and playback notifications.
Baseline Profile for faster startup and smoother scrolling.
Changes and removals
The 0.6.2 system localisation mechanism is removed: UI strings are inlined and values/strings.xml translations no longer follow the system locale.
Aligned with the contributor branch, reverting parts of 0.6.1 / 0.6.2 (artist-screen redesign, cross-provider matching, encrypted-session fix, crash store, and more).
Lyrics and other core logic stay on the mainline implementation and keep the AutoSync integration.
Known issues
The three AutoSync models remain NOT_ACCEPTED_FOR_WORD_LEVEL_ACCURACY; word-level accuracy has not been accepted against a human reference.
Japanese word timing is unavailable because the G2P dependency is not bundled.