Phrasync v0.1.0
Phrasync v0.1.0
This is the first public release: useful today, intentionally honest about what still needs engineering.
Highlights
- Two first-class project modes: Lyric Video for music and Subtitles for spoken audio or video.
- A video selected in Subtitles mode becomes the transcription source, preview footage, render background, and final audio source without duplicate setup.
- Hierarchical local Faster-Whisper transcription: ten-second language mapping, confidence and silence filtering, stable span smoothing, quiet-boundary snapping, language-locked span decoding, phrase-level language verification, selective corrective re-reading, adaptive fallback, and persisted gauntlet diagnostics.
- Auto follows language changes across a track; one code locks a language, comma-separated codes restrict the candidate set, and
singlekeeps one automatically detected language. - Language metadata is preserved on cues and returned as a dominant language, ordered language list, editable API-level span map, confidence, and diagnostics.
- Seven kinetic 2D typography presets with browser/Python timing parity.
- Seven separately named and tuned 3D typography personalities with independent world-space X/Y offsets.
- Grounded world-space 3D text: phrases are planted on the scene floor and the camera travels past them like other meshes.
- Odyssey 3D environments for Japan, Italy, China, and the USA.
- Manual rain, snow, fog, storms, falling leaves, dawn, day, sunset, night, and four seasons.
- Automatic journey mode crossfades daytime and seasonal colour while selecting deterministic weather from the playback clock.
- Image, video, and dynamic backgrounds; local project save/load; critic preflight; cancellable render jobs; H.264 MP4 download and postflight decoding.
Verified for this release
- 50 automated tests pass in the current release environment.
- A real short MP4 render passes with one video file used for both picture and audio in Subtitles mode.
- Chrome smoke coverage passes for every 2D/3D text composition state over Odyssey 3D, a 2D dynamic visual, image, and video modes; actual MP4 decoding is covered by the integration render above.
- 3D text remains world-planted and completes its fly-past at scene speeds of 20%, 100%, and 260%.
- Weather, daytime, seasons, automatic state changes, project-mode copy, and the 3D preset catalogue pass in the live browser with zero captured exceptions.
- Repeated 244.9-second
large-v3Auto runs map Italian, Portuguese, French, Spanish, English, and Japanese; recover the opening lyrics; preserve Japanese writing; follow late line-level switches; and produce no false outro after the final vocal. - Development and physical testing were performed on Windows 11, Ryzen 9 3900X, 32 GB RAM, and RTX 3090 24 GB. Linux and macOS launchers are included but not physically release-tested yet.
Known limits
- Odyssey Flat, Odyssey 3D, and 3D typography are preview-only in v0.1.0. MP4 preflight blocks them explicitly because the Python renderer does not yet reproduce WebGL output faithfully.
- Full song-length output has not yet been rendered. Current verification uses short engineering renders; complete music videos will be added after publication.
- Projects reference media in the local Phrasync workspace rather than embedding it into a portable archive.
- Browser video preview is intentionally limited to MP4 and WebM; convert MOV, MKV, AVI, and unusual codecs before import.
- The transcription workflow performs strongly on the tested songs, but expressive singing, unusual pronunciation, effects, and closely related languages can still need human review.
- There is no signed installer, automatic updater, or hosted web version.
Evaluated next-step technology
- Troika Three Text for SDF-quality, multilingual 3D typography.
- JASSUB/libass for professional ASS subtitle parity in browser preview.
- Three.js Sky for physically based atmosphere when deterministic export parity is ready.
- three.quarks only if future VFX outgrow the lightweight deterministic weather system.
Phrasync is created by jaydemks and distributed under Apache-2.0 with attribution preserved through NOTICE. Output belongs to its creator, subject to the rights they hold in the source music, text, footage, images, and fonts.