cat-shark
a8fe133aa4
解决转写漏句（有人说话但没识别出来）问题：silero VAD 对呻吟/轻语/BGM 混叠声学切段能力天然不足，把真话当非语音剔除（实测 savr-1054 全片仅 召回 115 条）。新增 decode_full 参数（默认 false 保持 VAD 现状）： - decode_full=true 时强制无 VAD 整段解码 + 跳过自动 VAD 分析，救回被 剔除的弱语音（savr-1054 全片 115 条 → 340 条） - 副作用是长时寒暄套话幻觉（おやすみなさい/ご視聴ありがとうございま した 等），whisper 转录后连带时间戳整条删除（clean_japanese_ha lllucinations），不留下 '-' 占位污染下游（占位会渲染进 ASS 成减号） - llm-translate 翻译后同样整条删除中文长时寒暄幻觉（clean_srt_text） - 短时（≤15s）相同词可能是剧情真实道晚安，保留（15s 阈值实测校准） - subtitle_cleanup 由 '-' 占位式改为整条删除式 + 剩余重编号，新增 JAPANESE_HALLUCINATION_TOKENS 词表 新增工作流 learn-translate（学习资料转译+翻译字幕）示范 decode_full 用法，并确立参数标注约定：params._note_<参数名> 存放设定理由与正反例、 _node_help 放节点参数手册（_ 前缀说明键，节点执行时忽略，零运行影响）。 调研记录见 docs/调研-whisper漏句与decode_full验证.md（A/B 实验、结论 修正与 5 个待决问题）。
166 lines
6.9 KiB
Python
166 lines
6.9 KiB
Python
"""字幕幻觉清洗测试（先红后绿）。
|
|
|
|
背景（实测 run 20260905115050）：修复时间对齐后，字幕仍残留四类问题，
|
|
其中"寒暄/收尾幻觉词"最具确定性、可规则化：
|
|
|
|
- 产物里 '晚安 / 感谢观看 / 感谢收看 / 感谢您的观看' 等固定套话出现 36 次；
|
|
- 其中 **33 条展示时长 = 30s（整块占满）**，明显是 ASR/LLM 对无内容段
|
|
的音量幻觉占位，与视频内容毫无关系；
|
|
- 仅 2 条时长 ~2s（如 720.00-722.00 '晚安'）可能是剧情里真的说了"晚安"，
|
|
属于真实内容，不应误删。
|
|
|
|
方案（用户 2026-09 确认改为**整条剔除**）：幻觉识别出后（展示时长 ≥ 阈值
|
|
且文本命中寒暄词表），应**连带时间戳把整条字幕 cue 删除**（剩余重新编号），
|
|
而不是替换成 '-' 占位——占位会一路流到 ASS 渲染成可见减号，处理位置绕且
|
|
不彻底。因此清洗统一为在幻觉产生处（whisper decode_full 转录后 / LLM 翻译
|
|
后）直接删除整条。
|
|
|
|
两类词表：
|
|
- HALLUCINATION_TOKENS：中文（LLM 翻译产物）；
|
|
- JAPANESE_HALLUCINATION_TOKENS：日文（whisper decode_full 直出）。
|
|
|
|
阈值从本次真实运行实测数据判定：30s 幻觉占位 vs 2s 真实词，分界明显，
|
|
本测试选 threshold=15s（≥15s 才删除；≤15s 保留）。
|
|
"""
|
|
|
|
from __future__ import annotations
|
|
|
|
import pytest
|
|
|
|
# 清洗逻辑来自生产模块 nodes/subtitle_cleanup.py（纯函数），测试只 import。
|
|
from nodes.subtitle_cleanup import (
|
|
DEFAULT_THRESHOLD_SECONDS,
|
|
HALLUCINATION_TOKENS,
|
|
clean_japanese_hallucinations,
|
|
clean_srt_text,
|
|
)
|
|
|
|
|
|
def test_remove_long_hallucination_whole_cue() -> None:
|
|
"""30s 的'晚安/感谢观看'（幻觉占位）须整条删除：序号+时间轴+文本都消失。"""
|
|
srt = (
|
|
"1\n00:00:00,000 --> 00:00:02,000\n真实内容\n\n"
|
|
"2\n00:01:00,000 --> 00:01:30,000\n晚安\n\n"
|
|
"3\n00:02:00,000 --> 00:02:30,000\n感谢您的观看\n\n"
|
|
"4\n00:03:00,000 --> 00:03:02,000\n继续真实\n\n"
|
|
)
|
|
out = clean_srt_text(srt)
|
|
# 幻觉条目连带时间戳整条消失。
|
|
assert "00:01:00,000 --> 00:01:30,000" not in out
|
|
assert "晚安" not in out
|
|
assert "00:02:00,000 --> 00:02:30,000" not in out
|
|
assert "感谢您的观看" not in out
|
|
# 真实条目保留且序号重新连续编号（原 1、4 -> 新 1、2）。
|
|
assert out.startswith("1\n00:00:00,000 --> 00:00:02,000\n真实内容")
|
|
assert "2\n00:03:00,000 --> 00:03:02,000\n继续真实" in out
|
|
|
|
|
|
def test_remove_short_hallucination_preserved() -> None:
|
|
"""2s 的'晚安'（剧情真实道晚安）必须保留，不误删。"""
|
|
srt = (
|
|
"1\n00:12:00,000 --> 00:12:02,000\n晚安\n\n"
|
|
"2\n00:12:03,000 --> 00:12:06,000\n明天见\n\n"
|
|
)
|
|
out = clean_srt_text(srt)
|
|
assert "晚安" in out
|
|
assert "明天见" in out
|
|
assert out.count("-->") == 2
|
|
|
|
|
|
def test_remove_non_hallucination_always_preserved() -> None:
|
|
"""普通内容（即使很长）绝不能被当成寒暄幻觉处理。"""
|
|
srt = (
|
|
"1\n00:00:00,000 --> 00:00:25,000\n"
|
|
"今天我将为您提供精神调适服务\n\n"
|
|
)
|
|
out = clean_srt_text(srt)
|
|
assert "精神调适" in out
|
|
|
|
|
|
def test_remove_threshold_boundary() -> None:
|
|
"""阈值边界：恰好 ≥ 阈值才删除；< 阈值保留。"""
|
|
srt_ge = "1\n00:00:00,000 --> 00:00:15,000\n晚安\n\n"
|
|
assert "晚安" not in clean_srt_text(srt_ge)
|
|
srt_lt = "1\n00:00:00,000 --> 00:00:14,990\n晚安\n\n"
|
|
assert "晚安" in clean_srt_text(srt_lt)
|
|
|
|
|
|
def test_remove_resequences_numbers() -> None:
|
|
"""删除中间 cue 后，剩余条目序号从 1 连续递增（合法 SRT）。"""
|
|
srt = (
|
|
"1\n00:00:00,000 --> 00:00:01,000\n甲\n\n"
|
|
"2\n00:01:00,000 --> 00:01:30,000\n晚安\n\n" # 幻觉被删
|
|
"3\n00:02:00,000 --> 00:02:01,000\n乙\n\n"
|
|
"4\n00:03:00,000 --> 00:03:01,000\n丙\n\n"
|
|
)
|
|
out = clean_srt_text(srt)
|
|
lines = [l for l in out.splitlines() if l.strip()]
|
|
# 重新编号：序号应为 1,2,3 各一次。
|
|
import re
|
|
numbers = [int(l) for l in lines if re.fullmatch(r"\d+", l.strip())]
|
|
assert numbers == [1, 2, 3]
|
|
|
|
|
|
# ---------------------------------------------------------------------------
|
|
# 日语（ASR 直出）幻觉清洗 —— whisper 节点 decode_full 兜底用
|
|
# ---------------------------------------------------------------------------
|
|
|
|
|
|
def test_jp_long_hallucination_removed() -> None:
|
|
"""30s 的'おやすみなさい/ご視聴ありがとうございました'（无语音段幻觉）整条删除。"""
|
|
srt = (
|
|
"1\n00:00:00,000 --> 00:00:02,000\n気持ちいい\n\n"
|
|
"2\n00:00:10,000 --> 00:00:40,000\nおやすみなさい\n\n"
|
|
"3\n00:00:41,000 --> 00:01:11,000\nご視聴ありがとうございました\n\n"
|
|
"4\n00:01:12,000 --> 00:01:14,000\nまた明日ね\n\n"
|
|
)
|
|
out = clean_japanese_hallucinations(srt)
|
|
assert "おやすみなさい" not in out
|
|
assert "ご視聴ありがとうございました" not in out
|
|
assert "00:00:10,000 --> 00:00:40,000" not in out
|
|
assert "気持ちいい" in out
|
|
assert "また明日ね" in out
|
|
|
|
|
|
def test_jp_short_hallucination_preserved() -> None:
|
|
"""2s 的'おやすみなさい'（剧情真实道晚安）须保留，不误删。"""
|
|
srt = (
|
|
"1\n00:12:00,000 --> 00:12:02,000\nおやすみなさい\n\n"
|
|
"2\n00:12:03,000 --> 00:12:06,000\nまた明日ね\n\n"
|
|
)
|
|
out = clean_japanese_hallucinations(srt)
|
|
assert "おやすみなさい" in out
|
|
assert "また明日ね" in out
|
|
|
|
|
|
def test_jp_non_hallucination_always_preserved() -> None:
|
|
"""普通长句（即使很长）绝不能被当成日语幻觉处理。"""
|
|
srt = (
|
|
"1\n00:00:00,000 --> 00:00:25,000\n"
|
|
"今日はお客様のために精神整備を務めさせていただきます\n\n"
|
|
)
|
|
out = clean_japanese_hallucinations(srt)
|
|
assert "精神整備" in out
|
|
|
|
|
|
def test_jp_threshold_15s() -> None:
|
|
"""日语清洗同样遵守 15s 时长阈值：15s 恰好删除，14.99s 保留。"""
|
|
srt = "1\n00:00:00,000 --> 00:00:15,000\nおやすみなさい\n\n"
|
|
assert "おやすみなさい" not in clean_japanese_hallucinations(srt)
|
|
srt2 = "1\n00:00:00,000 --> 00:00:14,990\nおやすみなさい\n\n"
|
|
assert "おやすみなさい" in clean_japanese_hallucinations(srt2)
|
|
|
|
|
|
@pytest.mark.integration
|
|
def test_jp_middle_removal_resequences() -> None:
|
|
"""删除中间日语幻觉后剩余条目重编号且文本/时间正确对应。"""
|
|
srt = (
|
|
"1\n00:00:00,000 --> 00:00:02,000\nあ\n\n"
|
|
"2\n00:00:10,000 --> 00:00:40,000\nおやすみなさい\n\n" # 删
|
|
"3\n00:00:41,000 --> 00:00:43,000\nい\n\n"
|
|
)
|
|
out = clean_japanese_hallucinations(srt)
|
|
assert out.count("-->") == 2
|
|
assert out.startswith("1\n00:00:00,000 --> 00:00:02,000\nあ")
|
|
assert "2\n00:00:41,000 --> 00:00:43,000\nい" in out
|