4.0 KiB
4.0 KiB
TalkEdit — Tech Stack, Tools, and Features
This document summarizes the chosen technology, tooling, the full feature set, recommended additions, and items on the back burner.
Overview
- Goal: Offline, local text-based audio/video editor (Descript-style) focused on spoken-word creators (podcasters, YouTubers). Fast, privacy-first, single-file installer.
Tech Stack
- Frontend: React 19 + Vite + TypeScript + Tailwind CSS + Zustand (with zundo undo/redo) + Virtuoso (virtualized transcript)
- Backend: Tauri 2.0 (Rust) for file I/O, licensing, licensing crypto (Ed25519), model management, error logging
- Transcription: Python faster-whisper with WhisperX for word-level alignment. Models downloaded on demand.
- Audio/Video Processing: FFmpeg invoked from Rust via Python scripts (video_editor.py, audio_cleaner.py, caption_generator.py)
- AI: Ollama, OpenAI, Claude through Python ai_provider.py. Bundled Qwen3 LLM planned.
- State: Zustand (in-frontend store) + zundo middleware for undo/redo history
- Packaging: Tauri
tauri buildfor cross-platform installers
Developer Tools
- Rust toolchain (cargo, rustc)
- Node.js + npm for frontend
- Python 3.11+ (faster-whisper, WhisperX, AI providers)
- FFmpeg binaries (platform-specific; bundled or downloaded at install)
- Build/test: Tauri CLI, Vite dev server
- Testing: Vitest (frontend), cargo test (Rust), pytest (Python)
- CI: GitHub Actions (Rust clippy/test, Frontend tsc/vitest, Python pytest)
Implemented Features
- 1. Media import via file dialog (audio/video auto audio-extract)
- 2. One-click local transcription with model selector (tiny/base → larger models) and model-size chooser
- 3. Scrollable, Google-Doc-style transcript editor (Virtuoso virtualized)
- Click word → seek video/audio
- Select words → cut corresponding media segment (smart 150–250ms fades)
- 4. Smart Cleanup
- Filler word removal (configurable list per-project)
- Silence trimming
- 5. Audio Polish chain (FFmpeg): normalize, compression, noise reduction
- 6. Preview with synced playback, undo/redo (zundo), project save/load
- 7. Export MP4/audio with SRT/VTT/ASS captions (speaker-labeled)
- 8. Speaker diarization
- 9. Custom filler lists per-project
- 10. Background music with auto-ducking
- 11. Append clips (concatenation)
- 12. Settings: AI provider config (Ollama, OpenAI, Claude)
- 13. Keyboard shortcuts with custom remapping
- 14. Help panel + cheatsheet
- 15. 7-day licensing with Ed25519-signed license keys
Recommended Additions (near-term, high ROI)
- Local GPU/CPU detection & recommended model/settings UI
- Per-project incremental transcription: re-run only edited segments
- "Preview cleaning" dry-run that highlights candidate removals before applying
- Export size/time estimator and suggested export presets
- Accessibility export presets (podcast vs YouTube presets)
- Bundled Qwen3 LLM for offline AI features
Remove / Defer (Back Burner)
These broaden scope or add legal/privacy surface — defer for now.
- Voice cloning / TTS: DEFER
- Multi-track, full timeline NLE features: DEFER
- Real-time collaboration / cloud sync: DEFER
- Built-in cloud processing by default: DEFER (make optional add-on later)
Risks & Mitigations
- Large model sizes: don't bundle large models; download on-demand and document storage location.
- Timestamp accuracy: WhisperX word-level alignment + manual per-segment re-run available.
- FFmpeg packaging/licensing: ship platform-specific binaries or use Tauri bundling guidance; document license compliance.
Prioritized Quick Wins
- Per-project incremental transcription
- "Preview cleaning" dry-run UI
- Export presets (podcast vs YouTube)
Next Steps for Implementation
- Bundle Qwen3 LLM for offline AI processing.
- Implement incremental transcription to speed up re-editing workflows.
- Add export presets and size estimation.
- Improve GPU/CPU detection and model recommendations.
Generated to capture tech, tools, implemented features, and the recommended add/remove/defer list.