todo: MoneyPrinterTurbo review and the OpenVoice feature gaps

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0182DhR1eFNPQyGmDrpHbedY

todo.md

@@ -19,3 +19,50 @@ For **unpixelate/upscale**, alternatives to Real-ESRGAN could include ESRGAN var
 The API should accept a `model` or `backend` parameter so the caller can select which implementation to use. Missing models should fail gracefully with a clear error rather than silently falling back. The web UI should expose the available options (discovered at runtime based on what's installed/downloaded) as a selector in the relevant panels.
 
 Read the existing implementations before proposing changes to understand the current structure, file layout, and how the UI communicates with the backend.
+
+---
+
+# TODO: review MoneyPrinterTurbo for ideas worth borrowing
+
+https://github.com/harry0703/MoneyPrinterTurbo/blob/main/README-en.md
+
+An automated short-video generation stack (script → stock/generated footage →
+subtitles → TTS → assembled clip). Worth reading against what we already have in
+`tools/videogen.py`, `tools/video_editor.py` and `tools/character_studio.py`:
+where does it do something we don't, and is any of it worth lifting? Points of
+comparison: how it plans a video from a single prompt, how it sources and cuts
+footage, subtitle generation/burn-in, the TTS/voice pipeline, and how its web UI
+is organised. Not yet evaluated — this entry is only a reminder to look.
+
+# TODO: OpenVoice features missing from our voice cloning
+
+https://github.com/myshell-ai/OpenVoice (MIT). We clone with F5-TTS
+(`codai/api/voice_clone.py`, in-context: reference audio + its transcript) and
+convert with seed-vc (`/v1/audio/convert`). OpenVoice splits the job — a base TTS
+speaks the text, then a tone-colour converter re-paints the timbre — which buys
+four things we lack, in rough order of value:
+
+1. **Transcript-free cloning.** `extract_se()` works from audio alone. Today a
+   profile without a transcript cannot clone at all (`video.py` falls back to a
+   generic voice), and our Whisper `base` fallback is the least accurate tier — a
+   wrong `ref_text` visibly degrades F5. Store a tone-colour embedding on the
+   profile at extract time and reuse it per line.
+2. **Cross-lingual voice.** The base TTS owns the language, so the accent doesn't
+   follow the reference. `tools/video_dubber.py` currently translates the text but
+   keeps the reference speaker's accent.
+3. **Multi-clip embedding averaging.** `extract_se()` averages several references;
+   we store exactly one WAV per profile. Cheapest quality win available, and it
+   fits the existing `PATCH /v1/audio/voices/{name}` shape.
+4. **Watermarking.** `add_watermark()` / `detect_watermark()` on synthesised
+   audio. We generate cloned speech of real people at scale and mark none of it.
+
+Integration shape: OpenVoice pins older torch/librosa, so it belongs in an
+isolated venv worker like pyannote / NeMo-Canary / H3 —
+`requirements-openvoice.txt` + `tools/openvoice_service.py` +
+`codai/api/openvoice_worker.py`. Then `/v1/audio/clone` gains an `engine` field
+(`f5` default, `openvoice`), auto-selecting OpenVoice when a profile has no
+transcript or the target language differs from the reference. Its converter sits
+beside seed-vc; both are audio→audio timbre transfer.
+
+Already done (2026-09-13, commit 45cc730): F5-TTS engine caching and reference
+trimming to 15 s — those were our own bugs, not missing OpenVoice features.