v0.1.0 — First test release

Reels Caption App — v0.1.0 (first test release)

A local-first desktop app that captions, styles, and voices short-form video —
transcription, styling, burning, and voiceover generation all run on your own
machine. The only network calls this app makes are Firebase Auth (sign-in) and,
optionally, fetching models on first use — no cloud transcription, no cloud TTS,
no per-use cost.

Captioning

• Automatic transcription the moment a video loads — no manual "run" step.
  English via NVIDIA Parakeet, Tamil via a dedicated Indic Whisper fine-tune,
  with word-level timestamps for precise, per-word caption timing.
• Automatic language detection — no language picker anywhere in the app.
  Speech-to-text detects the spoken language itself; text-to-speech detects the
  script's language from the text you type.
• Editable transcript — click any word to seek the video there, edit or
  delete words inline before burning.
• Five curated caption styles (Cascade Bold, Clean Classic, Boxed Minimal,
  Neon Pop, Highlight Emphasis) plus full manual control over font, size,
  color, position, background box, letter spacing, text case, and nine
  animation styles (karaoke fill, pop, bounce, typewriter, per-word highlight,
  slide-in, zoom, fade, or none).
• Burn & Export — styled captions are burned directly into the video via
  ffmpeg + libass, with automatic parallel segmented encoding on longer clips
  and hardware-encoder acceleration (NVENC/QSV/AMF, falling back to software)
  when your machine has one.

Editing

• Automatic silence & filler-word removal — trims dead air and "um"/"uh"
  in one pass, snapped to natural gaps so no caption gets cut mid-sentence.
• Speaker diarization — automatically detects distinct speakers and colors
  each one's captions differently.
• Prosody-driven emphasis — words spoken with more vocal energy render
  larger in Cascade mode, driven by the actual audio, not guesswork.
• Loudness normalization and background-music ducking (auto-quiets a
  music bed under speech).

Voice-over

• Write a script or upload a recording to replace a clip's audio (and its
  captions) entirely — generated locally via Piper (English) or MMS-TTS
  (Tamil).
• Voice cloning — the generated voiceover is automatically reshaped to
  match the original speaker's actual voice (via OpenVoice V2), not just a
  generic default. Falls back to a gender-matched preset voice when cloning
  isn't possible (e.g. a silent source clip), and always to a sensible default
  otherwise — never blocks voiceover generation.
• Automatic mouth-movement sync — timing-corrects the voiceover against
  the visible speaker's mouth movement in the loaded video.
• Emotion-aware delivery (optional) — classifies the script's tone and
  adjusts pacing/pitch to match, without needing a real performance.

Ideas

• AI-generated titles, descriptions, hashtags, and emoji from the
  transcript, via a small local LLM (Qwen2.5-0.5B-Instruct through
  llama.cpp) — no external API, no cost per generation.

What's included in this installer vs. what needs one-time setup

The .msi bundles everything needed for captioning, styling, burning,
silence removal, and title/hashtag generation out of the box — ffmpeg, the
local LLM, and the Tamil font are all included.

Transcription, voiceover, voice cloning, mouth-sync, and diarization/prosody
need a one-time local setup (Python/conda environments + a couple of
model downloads) — these are large, per-feature models deliberately kept out
of the installer. After installing, run:

npm run fetch-resources

from a clone of the repo (see the README) to pull the rest in one shot, then
follow the README's conda environment setup steps (sections 2, 2.4, 2.6, 2.7).
Installing the .msi alone will not give you working transcription —
please read the README before assuming something's broken.

Platform

Windows x64 only in this release (.msi, WiX-based installer). macOS and
Linux builds aren't cross-compiled from this release yet — see the README's
"Build & distribute" section for building those natively.

Known limitations

• Sign-in (Firebase Auth) is required before using the app at all.
• This is an early test build — please file issues for anything that breaks.