Integração: falantes + dublagem + vídeo final

- editor_api: importa dub e speakers juntos.
- voice_chunks (falantes) vira a migração v15, depois da v14 do vídeo final (antes era criada fora das migrações).
- dub.voice_for: qualquer voz que não é o streamer (other1, other2…) usa a segunda voz da dublagem.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
Mateuus
2026-11-18 03:59:44 -03:00
co-authored by Claude Opus 5.5
24 changed files with 1671 additions and 129 deletions
+11 -13
View File
@@ -21,7 +21,7 @@ from fastapi import FastAPI, Request
from fastapi.responses import JSONResponse
from pydantic import BaseModel, Field
from . import dub, editor, editor_render, layout, media
from . import dub, editor, editor_render, layout, media, speakers
from .config import REPO_ROOT
from .db import loads, now
from .judge import JudgeError, _chat_block, _transcript
@@ -170,9 +170,7 @@ def auto_for(engine, clip: dict, session: dict) -> dict:
"""Projeto que a IA entrega: o que o pipeline já sabe (análise, Whisper, volume, configuração do corte)."""
judge = _judge(clip)
lo, hi = editor.default_window(clip, session["duration_s"])
rows = engine.db.all("SELECT words_json FROM segments WHERE session_id = ? AND end_s > ? AND start_s < ? ORDER BY idx",
clip["session_id"], lo, hi)
words = [Word(a, b, c) for r in rows for a, b, c in loads(r["words_json"], [])]
words = engine.words_between(clip["session_id"], lo, hi) # já com quem fala cada palavra (FALANTES.md)
loudness = engine.load_timeline(clip["session_id"], max(0.0, lo - 5))["loudness"]
return editor.auto_project(clip, judge.get("analysis"), words, loudness, session["duration_s"],
engine.intel.settings_for(session), model=judge.get("model") or "", generated_at=now())
@@ -259,9 +257,7 @@ def register(app: FastAPI, engine, clip_view, get_clip) -> None:
return round(lo, 3), round(hi, 3)
def _words(session_id: int, lo: float, hi: float) -> list[Word]:
rows = db.all("SELECT words_json FROM segments WHERE session_id = ? AND end_s > ? AND start_s < ? ORDER BY idx",
session_id, lo, hi)
return [Word(a, b, c) for r in rows for a, b, c in loads(r["words_json"], []) if lo <= a <= hi]
return [w for w in engine.words_between(session_id, lo, hi) if lo <= w.start <= hi]
def _cached_proxy(clip: dict, lo: float, hi: float) -> dict | None:
f = edit_dir(engine, clip["session_id"], clip["id"]) / f"proxy-{_window_key(lo, hi)}.mp4"
@@ -301,7 +297,7 @@ def register(app: FastAPI, engine, clip_view, get_clip) -> None:
"origin": row["origin"] if row else "ai", "aiProject": ai_project,
"exportedAt": row["exported_at"] if row else None, "cover": row["cover"] if row else None,
"window": {"lo": w_lo, "hi": w_hi}, "sessionDurationS": session["duration_s"],
"words": [{"s": w.start, "e": w.end, "w": w.word} for w in _words(clip["session_id"], w_lo, w_hi)],
"words": [editor.word_item(w) for w in _words(clip["session_id"], w_lo, w_hi)],
"chat": chat, "chatPerS": per_s, "loudnessDb": loud, "analysis": analysis,
"settings": engine.intel.settings_for(session), "credit": engine.channel_label(session),
"presets": [{"id": k, "label": editor.PRESET_LABEL[k], "style": editor.PRESET_STYLE[k]} for k in editor.CAPTION_PRESETS],
@@ -352,8 +348,7 @@ def register(app: FastAPI, engine, clip_view, get_clip) -> None:
editor_render.make_proxy(ffmpeg, session_segments(engine, clip["session_id"]), lo, hi, folder / name)
log.info("proxy do corte %s (%.0f s de vídeo) em %.1f s", clip_id, hi - lo, time.monotonic() - t)
_clean_old(folder, "proxy", name)
return _cached_proxy(clip, lo, hi) | {"words": [{"s": w.start, "e": w.end, "w": w.word}
for w in _words(clip["session_id"], lo, hi)]}
return _cached_proxy(clip, lo, hi) | {"words": [editor.word_item(w) for w in _words(clip["session_id"], lo, hi)]}
@app.get("/v1/clips/{clip_id}/waveform")
def get_waveform(clip_id: int, lo: float, hi: float):
@@ -441,8 +436,11 @@ def register(app: FastAPI, engine, clip_view, get_clip) -> None:
words = engine.voice.transcribe(wav, f"clipper-edit-{clip_id}-{uuid.uuid4().hex[:12]}")
except Exception as exc: # noqa: BLE001 - erro do serviço de voz vira mensagem
raise editor.EditError(f"transcrição falhou: {str(exc)[:200]}") from exc
# Quem fala: o mesmo das palavras antigas no mesmo instante (a re-transcrição não sabe o falante).
fresh = speakers.transfer([Word(round(w.start + a, 3), round(w.end + a, 3), w.word) for w in words],
_words(clip["session_id"], a - 1, b + 1))
return {"startS": a, "endS": b, "seconds": round(time.monotonic() - t, 1),
"words": [{"s": round(w.start + a, 3), "e": round(w.end + a, 3), "w": w.word} for w in words]}
"words": [editor.word_item(w) for w in fresh]}
@app.post("/v1/clips/{clip_id}/tts")
def tts(clip_id: int, body: TtsIn):
@@ -485,12 +483,12 @@ def register(app: FastAPI, engine, clip_view, get_clip) -> None:
project = validate_for(engine, clip, session, data)
lo, hi = project["source"]["lo"], project["source"]["hi"]
tl = engine.load_timeline(clip["session_id"], max(0.0, lo - 5))
words = [Word(w["s"], w["e"], w["w"]) for w in project["captions"]["words"]] or tl["words"]
words = [Word(w["s"], w["e"], w["w"], w.get("sp") or "") for w in project["captions"]["words"]] or tl["words"]
view = clip_view(clip)
analysis = (view["judge"] or {}).get("analysis") if isinstance(view["judge"], dict) else None
cfg = editor.effective_settings(engine.intel.settings_for(session), project)
prompt = editor.suggest_prompt(engine.streamer_name(session), clip["title"], project, tl["chat"], analysis,
_transcript(words, lo, hi), _chat_block(tl["chat"], lo, hi, clip["peak_s"]), cfg)
_transcript(words, lo, hi, streamer=engine.streamer_name(session)), _chat_block(tl["chat"], lo, hi, clip["peak_s"]), cfg)
t = time.monotonic()
try:
raw = engine.judge.ask(editor.SUGGEST_SYSTEM, prompt, editor.SUGGEST_SCHEMA, [])