docs: face/voice collection spec + consent/legal

docs/footage-collection-spec.md

@@ -0,0 +1,66 @@
+# Footage & Voice Collection Spec — Target Person
+
+The avatar's identity is locked by two zero-shot assets. **No model training is
+required** — you replace two files referenced in `config/meeting_avatar.yaml`.
+
+## 1. Face asset — avatar base video (for MuseTalk)
+
+Replace `AvatarMusetalk.avatar_video_path` with your clip.
+
+| Requirement | Spec |
+|---|---|
+| Length | 1–3 minutes (it will loop) |
+| Resolution | 1080p preferred, 720p minimum |
+| Frame rate | 24–30 fps |
+| Framing | Head + shoulders, face filling ~40–60% of frame height |
+| Camera | Fully static (tripod), no zoom, no cuts |
+| Pose | Frontal (±15°), eyes mostly toward camera |
+| Mouth | Gently closed or minimal movement; occasional natural blinks |
+| Lighting | Even, frontal, no strong shadows or backlight |
+| Background | Plain/static preferred; no other faces in frame |
+| Format | MP4 (H.264), constant frame rate |
+
+**How to shoot it:** have the person sit still and "listen" — imagine being on a
+call while someone else talks. Subtle natural motion is good; talking with wide
+mouth movement is bad (it becomes the loop the lip-syncer has to overwrite).
+
+**Validation:** play the clip muted on loop for 2 minutes. If nothing looks
+robotic or jarring at the loop point, it is usable. Trim to hide the cut.
+
+## 2. Voice asset — reference audio (for CosyVoice2 zero-shot cloning)
+
+Replace `CosyVoice.ref_audio_path` and `ref_audio_text`.
+
+| Requirement | Spec |
+|---|---|
+| Length | 5–10 seconds (hard cap ~10 s) |
+| Format | WAV, 24 kHz or higher, mono |
+| Room | Quiet, minimal reverb (closet/blankets work), no music/noise |
+| Content | Natural conversational sentences — the person's typical speaking tone |
+| Transcript | Exact verbatim text, punctuation included → goes in `ref_audio_text` |
+
+**Validation:** the clone is only as good as this clip. Record 5 takes, pick the
+cleanest. Avoid whispering, shouting, or exaggerated delivery.
+
+## 3. Optional — LoRA upgrade path (photoreal streaming, later)
+
+If you later want LiveAvatar-class photoreal streaming instead of MuseTalk
+footage lip-sync: collect **15–20 clips × 30–60 s** of the person talking
+(diverse expressions/head angles, same quality bar as above) — the per-person
+video-LoRA recipe (6k–9k steps on a Wan-S2V-class base) trains from roughly
+that. This is a separate workstream; the current stack does not need it.
+
+## 4. Consent, disclosure, and platform rules (do not skip)
+
+- Obtain **written consent** from the person whose face/voice is cloned, and
+  keep it on file. You need it for both legal and platform-policy reasons.
+- **Disclose that the participant is an AI avatar** in meetings. Several
+  jurisdictions require it for synthetic media of a real person; some meeting
+  platforms (Zoom/Meet/Teams) have AI-participant policies that require host
+  consent. Violations can get accounts banned.
+- Impersonating someone without consent can violate right-of-publicity and
+  deepfake laws. Internal/prank use of your own likeness is the safe zone;
+  customer-facing use needs a lawyer's review.
+- Persona and Veriff (if those were meant in your platform list) are
+  identity-verification (KYC) providers, not meeting platforms — integrating
+  them would be for verifying the avatar's owner, a separate workstream.