nex_animation/backend/tests/test_speaker_voice_binding.py

194 lines
9.6 KiB
Python
Raw Permalink Blame History

This file contains ambiguous Unicode characters!

This file contains ambiguous Unicode characters that may be confused with others in your current locale. If your use case is intentional and legitimate, you can safely ignore this warning. Use the Escape button to highlight these characters.

"""回归用例:台词必须配给「真的在说话的那个人」。
项目 9《自检-异步调度验证》现场缺陷(用户红线「配音接错人」):
镜头 02 的画面是「俊杰坐在服务台后,抬手示意紫小慧不用担心」,台词明显是俊杰说的,
但俊杰是脚本**导入之后**才补进项目的人物——导入时匹配不到说话人就把
`speaker_character_id` 留空,旧逻辑随即退回「出镜名单第一位」(紫小慧),
于是男主整段台词配成了女声,而且这条女声线还会被固化写回镜头,
用户后来在资产库里新配的男声一次都没被用上(事件 3328/3398 实测)。
"""
from __future__ import annotations
from sqlalchemy import select
from app.constants import AssetKind
from app.models import EventLog, Project
from app.services import audio, storage
SCRIPT = """# 镜头 1:稳住
画面: 俊杰坐在服务台后,微微抬手示意紫小慧不用担心
景别: 中景
运镜: 固定
时长: 2
台词: 别慌,数据都在机房云桌面里。
说话人: 俊杰
出镜角色: 紫小慧
场景: 服务大厅
# 镜头 2:崩了
画面: 紫小慧抱着一摞文件冲进服务台,俊杰在旁边让开路
景别: 中景
运镜: 跟镜
时长: 2
台词: 完了完了,三号服务器全红了!
说话人: 紫小慧
出镜角色: 紫小慧
场景: 服务大厅
"""
def _shots(db, pid):
project = db.get(Project, pid)
return sorted(project.shots or [], key=lambda row: (row.order_index, row.id))
def _cast_project(api, db):
"""照搬项目 9 的真实形状:导入时项目里只有女主角,俊杰是之后才补的人物。"""
pid = api.project(name="自检-配音接错人")
female = api.post(f"/api/projects/{pid}/characters",
json={"name": "紫小慧", "role": "主角", "gender": "female",
"appearance": "黑色长发,紫色工装外套",
"persona_prompt": "急性子的运维工程师",
"voice_profile": "明亮女声"})["result"]["id"]
female_voice = api.post(f"/api/projects/{pid}/voices",
json={"name": "紫小慧声音", "provider": "mock",
"provider_voice_id": "v-female", "gender": "female",
"is_default": True})["result"]["id"]
api.patch(f"/api/projects/{pid}/characters/{female}",
json={"default_voice_id": female_voice})
api.post(f"/api/projects/{pid}/scenes",
json={"name": "服务大厅", "description": "公司一层服务台", "prompt": "服务大厅"})
api.post(f"/api/projects/{pid}/import",
data={"script_text": SCRIPT, "default_duration": 2})
# 导入之后用户才补的男主 + 他的男声(脚本里的说话人已经永久丢了)
male = api.post(f"/api/projects/{pid}/characters",
json={"name": "俊杰", "role": "主角", "gender": "male",
"appearance": "短发,深蓝西装", "persona_prompt": "稳重的IT工程师",
"voice_profile": "低沉男声"})["result"]["id"]
male_voice = api.post(f"/api/projects/{pid}/voices",
json={"name": "俊杰声音", "provider": "mock",
"provider_voice_id": "v-male", "gender": "male"})["result"]["id"]
api.patch(f"/api/projects/{pid}/characters/{male}", json={"default_voice_id": male_voice})
db.expire_all()
rows = _shots(db, pid)
# 脚本里写了「说话人: 俊杰」,可当时项目里还没有这个人——导入直接留空,
# 原始名字就此丢失,之后无论谁来配音都只能瞎猜(这正是接错人的源头)。
assert rows[0].speaker_character_id is None, rows[0].speaker_character_id
assert rows[1].speaker_character_id == female
assert all([row.id for row in (row.characters or [])] == [female] for row in rows)
return {"project_id": pid, "female_id": female, "male_id": male,
"female_voice": female_voice, "male_voice": male_voice}
def _actions(db, pid, shot_id=None):
stmt = select(EventLog).where(EventLog.project_id == pid)
rows = db.scalars(stmt).all()
if shot_id is not None:
rows = [row for row in rows if row.shot_id == shot_id]
return [row for row in rows]
def test_speaker_is_the_character_named_first_in_the_shot(api, db):
"""台词挂在谁身上:显式说话人 → 画面描述里最早点名的人 → 出镜名单第一位。"""
ctx = _cast_project(api, db)
male_shot, female_shot = _shots(db, ctx["project_id"])
assert audio.speaker_character(db, male_shot).name == "俊杰", \
"画面第一句就是俊杰的动作,台词不该算到紫小慧头上"
assert audio.speaker_gender(db, male_shot) == "male"
assert audio.speaker_character(db, female_shot).name == "紫小慧"
assert audio.speaker_gender(db, female_shot) == "female"
def test_speaker_follows_the_earlier_name_when_the_order_swaps(api, db):
"""同一段话里两个人都被点名时,认「先被点名」的那个,不能永远认女一号。"""
ctx = _cast_project(api, db)
shot = _shots(db, ctx["project_id"])[0]
shot.description = "紫小慧冲进来喊停,俊杰却已经把机房门关上了"
db.commit()
assert audio.speaker_character(db, shot).name == "紫小慧"
shot.description = "俊杰先把机房门关上,紫小慧被拦在外面"
db.commit()
assert audio.speaker_character(db, shot).name == "俊杰"
# 谁都没被点名时退回出镜名单第一位,保持旧口径不出新问题
shot.description = "服务大厅的灯一盏盏熄灭"
db.commit()
assert audio.speaker_character(db, shot).id == ctx["female_id"]
def test_male_dialogue_no_longer_gets_the_female_lead_voice(api, db):
"""男主的台词必须匹配到男声线,且不能被项目默认(女声)兜底走。"""
ctx = _cast_project(api, db)
project = db.get(Project, ctx["project_id"])
male_shot, female_shot = _shots(db, ctx["project_id"])
voice, info = audio.resolve_voice(db, project, male_shot)
assert voice.name == "俊杰声音", info
assert info["gender"] == "male" and info["gender_mismatch"] is False, info
assert "俊杰声音" in info["reason"], info["reason"]
other, other_info = audio.resolve_voice(db, project, female_shot)
assert other.name == "紫小慧声音", other_info
assert other_info["gender_mismatch"] is False
def test_generate_voiceover_records_the_inferred_speaker(api, db):
"""推断出来的说话人要写回镜头、挂进出镜名单,并在事件里说清楚——用户能一眼改。"""
ctx = _cast_project(api, db)
project = db.get(Project, ctx["project_id"])
shot = _shots(db, ctx["project_id"])[0]
result = audio.generate_voiceover(db, project, shot)
assert result["voice"] == "俊杰声音", result
db.expire_all()
shot = _shots(db, ctx["project_id"])[0]
assert shot.speaker_character_id == ctx["male_id"], "配音完却看不出是谁在说话"
assert ctx["male_id"] in {row.id for row in (shot.characters or [])}, \
"说话人必须进本镜出镜名单,否则定妆图/口型参照都找不到他"
assert storage.first_asset(db, "shot", shot.id, AssetKind.VOICEOVER) is not None
inferred = [row for row in _actions(db, ctx["project_id"], shot.id)
if row.action == "voiceover.speaker_inferred"]
assert len(inferred) == 1, [row.action for row in _actions(db, ctx["project_id"], shot.id)]
assert inferred[0].data_json["speaker"] == "俊杰", inferred[0].data_json
assert inferred[0].data_json["voice"] == "俊杰声音"
assert "俊杰声音" in inferred[0].message
# 性别不再错配:事件流里不该再出现「说话人是男声,声线却是女声」这种告警
assert all("却是女声" not in row.message for row in
_actions(db, ctx["project_id"], shot.id)
if row.action == "voiceover.voice_fallback")
def test_inference_is_idempotent_and_leaves_explicit_speakers_alone(api, db):
"""用户/脚本已经写明说话人时绝不猜,也不再产生推断事件。"""
ctx = _cast_project(api, db)
project = db.get(Project, ctx["project_id"])
shot = _shots(db, ctx["project_id"])[0]
shot.speaker_character_id = ctx["female_id"] # 编导明确说:这句是紫小慧说的
db.commit()
assert audio.inferred_speaker(db, shot) is None
result = audio.generate_voiceover(db, project, shot)
assert result["voice"] == "紫小慧声音", result
db.expire_all()
shot = _shots(db, ctx["project_id"])[0]
assert shot.speaker_character_id == ctx["female_id"], "显式说话人被推断覆盖了"
assert "voiceover.speaker_inferred" not in [row.action for row
in _actions(db, ctx["project_id"], shot.id)]
def test_no_inference_event_when_the_cast_first_already_speaks(api, db):
"""结论和旧口径一致时不该刷事件,免得把用户淹死在噪音里。"""
ctx = _cast_project(api, db)
project = db.get(Project, ctx["project_id"])
shot = _shots(db, ctx["project_id"])[1] # 紫小慧:既是出镜第一位也是说话人
shot.speaker_character_id = None # 历史数据里说话人丢了
db.commit()
assert audio.inferred_speaker(db, shot) is None
audio.generate_voiceover(db, project, shot)
db.expire_all()
shot = _shots(db, ctx["project_id"])[1]
assert shot.speaker_character_id is None
assert "voiceover.speaker_inferred" not in [row.action for row
in _actions(db, ctx["project_id"], shot.id)]