fix: whisper 重复伪影整条删除,避免翻译层空响应失败
whisper 在单个窗口内卡住重复时会把一个单元写满整条 cue(实测 30 秒整、重复 74–446 次,如 `チン`×111)。这类正文会让下游 LLM 跟着循环、把输出预算耗在思考上, 最终 `content` 返回空串并报 `translation alignment failed … Expecting value: line 1 column 1 (char 0)`;它也可能直接渲染成超长字幕行。 - `nodes/subtitle_cleanup.py` 新增 `remove_repetition_entries`:**纯模式判据、无字符 词表**——展示时长 ≥15s 且同一 1–6 字单元连续重复 ≥6 次且覆盖正文 ≥70% 的 cue 整条 删除;真实短促呻吟(`ぇ`×15、`ああああああ`)靠时长区分,零误删。 - `nodes/whisper.py` 在转写产出处应用该清洗(VAD 开关都生效,它与套话幻觉无关)。 - 真实数据验证:本地全部真实转写 3605 条 cue 只删 5 条 30 秒伪影;对照实验同批次 伪影截短后 5/5 成功、原样 1/5。
This commit is contained in:
@@ -8,6 +8,9 @@ whisper(日语链路)与 llm-translate(中文链路)复用,纯函数
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import json
|
||||
from pathlib import Path
|
||||
|
||||
from nodes.subtitle_cleanup import (
|
||||
DEFAULT_MOAN_MAX_CHARS,
|
||||
HALLUCINATION_TOKENS,
|
||||
@@ -15,6 +18,7 @@ from nodes.subtitle_cleanup import (
|
||||
clean_japanese_hallucinations,
|
||||
clean_srt_text,
|
||||
remove_hallucination_entries,
|
||||
remove_repetition_entries,
|
||||
remove_short_moan_entries,
|
||||
)
|
||||
from tests.shared.srt_entries import parse_srt_entries
|
||||
@@ -246,3 +250,83 @@ def test_multiline_moan_entry_removed_as_one_cue() -> None:
|
||||
# 验证结果:只剩第二条并重编号。
|
||||
assert [e["text"] for e in parse_srt_entries(cleaned)] == ["そこ"]
|
||||
assert cleaned.startswith("1\n")
|
||||
|
||||
|
||||
# 真实转写抽样(data/repetition_cues.json):30 秒窗口被同一单元填满的 whisper
|
||||
# 重复伪影 + 真实短呻吟 + 真实台词,用于"重复伪影"判据的正反例。
|
||||
_REPETITION_CUES = json.loads(
|
||||
(Path(__file__).parent / "data" / "repetition_cues.json").read_text(encoding="utf-8")
|
||||
)
|
||||
|
||||
|
||||
def test_remove_repetition_entries_deletes_whisper_loops_only() -> None:
|
||||
"""数据:真实 ASR 产物——3 条 30 秒重复伪影(重复 74/111/446 次)、
|
||||
3 条 2–3 秒真实呻吟(重复 6–10 次)、1 条正常台词。
|
||||
|
||||
过程:调用 remove_repetition_entries 清理整份 SRT。
|
||||
|
||||
验证:只删 30 秒伪影,真实呻吟与台词原样保留,剩余 cue 序号连续。
|
||||
"""
|
||||
artifacts = _REPETITION_CUES["artifacts"]
|
||||
moans = _REPETITION_CUES["moans"]
|
||||
normal = _REPETITION_CUES["normal"]
|
||||
srt = _srt(
|
||||
*[(c["start"], c["end"], c["text"]) for c in artifacts],
|
||||
*[(c["start"], c["end"], c["text"]) for c in moans],
|
||||
(normal["start"], normal["end"], normal["text"]),
|
||||
)
|
||||
|
||||
cleaned = remove_repetition_entries(srt)
|
||||
|
||||
for cue in artifacts:
|
||||
assert cue["text"] not in cleaned
|
||||
for cue in moans:
|
||||
assert cue["text"] in cleaned
|
||||
assert normal["text"] in cleaned
|
||||
entries = parse_srt_entries(cleaned)
|
||||
assert [e["text"] for e in entries] == [c["text"] for c in moans] + [normal["text"]]
|
||||
# 序号/时间轴重建后从 1 连续编号,不留空号(合法 SRT)。
|
||||
numbers = [line for line in cleaned.splitlines() if line.strip().isdigit()]
|
||||
assert numbers == [str(i) for i in range(1, len(entries) + 1)]
|
||||
|
||||
|
||||
def test_remove_repetition_entries_keeps_short_repeated_moan() -> None:
|
||||
"""数据:3 秒内重复 15 次的真实呻吟(时长不足阈值)。
|
||||
|
||||
过程:调用 remove_repetition_entries。
|
||||
|
||||
验证:保留——时长阈值是"窗口被填满"的判据,短促重复属真实发声。
|
||||
"""
|
||||
srt = _srt(("00:00:01,000", "00:00:04,200", "ぇ" * 15))
|
||||
|
||||
cleaned = remove_repetition_entries(srt)
|
||||
|
||||
assert "ぇ" * 15 in cleaned
|
||||
|
||||
|
||||
def test_remove_repetition_entries_keeps_mixed_long_line() -> None:
|
||||
"""数据:30 秒长条但正文以正常台词为主,只有少量重复。
|
||||
|
||||
过程:调用 remove_repetition_entries。
|
||||
|
||||
验证:保留——重复片段未占正文 70% 以上,不构成重复伪影。
|
||||
"""
|
||||
text = "そうですね、それでいいと思いますよ" * 3 + "ああ"
|
||||
srt = _srt(("00:00:01,000", "00:00:31,000", text))
|
||||
|
||||
cleaned = remove_repetition_entries(srt)
|
||||
|
||||
assert text in cleaned
|
||||
|
||||
|
||||
def test_remove_repetition_entries_can_be_disabled() -> None:
|
||||
"""数据:一条 30 秒重复伪影,阈值设为 0(关闭)。
|
||||
|
||||
过程:调用 remove_repetition_entries(threshold_seconds=0)。
|
||||
|
||||
验证:原样返回,便于按需走旧行为。
|
||||
"""
|
||||
artifact = _REPETITION_CUES["artifacts"][0]
|
||||
srt = _srt((artifact["start"], artifact["end"], artifact["text"]))
|
||||
|
||||
assert remove_repetition_entries(srt, threshold_seconds=0) == srt
|
||||
|
||||
Reference in New Issue
Block a user